I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
2025
,
pp.
3395
~
3403
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
33
95
-
3403
3395
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
Fin
e
-
t
u
n
in
g b
i
d
ir
e
c
t
io
n
al
e
n
c
od
e
r
r
e
p
r
e
se
n
t
a
t
io
n
s f
r
om
t
r
an
s
f
or
m
e
r
s
f
or
t
h
e
X
soc
ia
l
m
e
d
ia
p
e
r
son
al
ity d
e
t
e
c
t
io
n
S
e
lvi
F
i
t
r
ia
Kh
oe
r
u
n
n
is
a
1
,
B
ayu
S
u
r
ar
s
o
2
,
Re
t
n
o
Ku
s
u
m
an
in
gr
u
m
3
1
M
a
s
te
r
of
I
nf
or
ma
ti
on
S
ys
te
m
, S
c
hool
of
P
os
tg
r
a
dua
te
S
tu
di
e
s
, U
ni
ve
r
s
it
a
s
D
ip
one
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
2
D
e
pa
r
tm
e
nt
of
M
a
th
e
ma
ti
c
s
, F
a
c
ul
ty
of
S
c
ie
nc
e
a
nd
M
a
th
e
ma
t
ic
s
, U
ni
ve
r
s
it
a
s
D
ip
on
e
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
3
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
c
s
, F
a
c
ul
ty
of
S
c
ie
nc
e
a
nd M
a
th
e
m
a
ti
c
s
, U
ni
v
e
r
s
it
a
s
D
ip
one
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
M
a
y
18,
2024
R
e
vis
e
d
Apr
15,
2025
Ac
c
e
pted
J
un
8,
2025
U
n
d
ers
t
an
d
i
n
g
p
ers
o
n
al
i
t
y
t
ra
i
t
s
can
h
el
p
i
n
d
i
v
i
d
u
a
l
s
reach
t
h
e
i
r
fu
l
l
p
o
t
en
t
i
a
l
an
d
h
as
ap
p
l
i
cat
i
o
n
s
i
n
v
ari
o
u
s
fi
el
d
s
s
u
ch
as
recru
i
t
me
n
t
,
ad
v
er
t
i
s
i
n
g
,
an
d
mark
et
i
n
g
.
A
w
i
d
e
l
y
u
s
ed
t
o
o
l
fo
r
a
s
s
es
s
i
n
g
p
ers
o
n
a
l
i
t
y
i
s
My
ers
-
Br
i
g
g
s
t
y
p
e
i
n
d
i
ca
t
o
r
(MB
T
I).
Rec
en
t
a
d
v
a
n
ce
men
t
s
i
n
t
ech
n
o
l
o
g
y
h
av
e
al
l
o
w
ed
f
o
r
res
earch
o
n
h
o
w
p
ers
o
n
a
l
i
t
i
e
s
can
ch
an
g
e
b
a
s
ed
o
n
s
o
ci
a
l
med
i
a
u
s
e.
Prev
i
o
u
s
res
earch
u
s
ed
mach
i
n
e
l
ear
n
i
n
g
met
h
o
d
s
,
d
ee
p
l
earn
i
n
g
met
h
o
d
s
,
u
n
t
i
l
t
ran
s
fo
rmer
s
-
b
as
e
d
met
h
o
d
.
H
o
w
e
v
er,
t
h
es
e
p
r
ev
i
o
u
s
ap
p
r
o
ach
e
s
mu
s
t
b
e
rev
i
s
ed
t
o
req
u
i
re
ex
t
e
n
s
i
v
e
d
at
a
an
d
a
h
i
g
h
co
mp
u
t
a
t
i
o
n
a
l
l
o
a
d
.
A
l
t
h
o
u
g
h
t
ran
s
fo
rmer
-
b
as
e
d
met
h
o
d
s
l
i
k
e
b
i
d
i
r
ect
i
o
n
al
en
co
d
er
rep
re
s
en
t
at
i
o
n
s
fr
o
m
t
ra
n
s
f
o
rmers
(B
E
RT
)
e
x
cel
at
u
n
d
ers
t
an
d
i
n
g
co
n
t
ex
t
,
i
t
s
t
i
l
l
h
a
s
l
i
m
i
t
a
t
i
o
n
s
i
n
cap
t
u
r
i
n
g
w
o
rd
o
rd
er
an
d
s
t
y
l
i
s
t
i
c
v
ari
a
t
i
o
n
s
.
T
h
eref
o
re,
t
h
i
s
s
t
u
d
y
p
r
o
p
o
s
e
d
i
n
t
e
g
rat
i
n
g
fi
n
e
-
t
u
n
i
n
g
B
E
RT
w
i
t
h
recu
rren
t
n
eu
ra
l
n
e
t
w
o
rk
s
(RN
N
s
)
co
n
s
i
s
t
i
n
g
o
f
v
a
n
i
l
l
a
RN
N
,
l
o
n
g
s
h
o
r
t
-
t
erm
memo
ry
(L
ST
M),
an
d
g
at
e
d
recu
rren
t
u
n
i
t
(G
RU
).
T
h
i
s
s
t
u
d
y
al
s
o
u
s
es
a
BE
RT
b
a
s
e
fu
l
l
y
c
o
n
n
ect
e
d
l
a
y
er
a
s
a
c
o
mp
ar
i
s
o
n
.
T
h
e
res
u
l
t
s
s
h
o
w
t
h
at
t
h
e
B
E
RT
b
as
e
fu
l
l
y
co
n
n
ec
t
ed
l
a
y
er
a
p
p
r
o
ach
s
t
r
at
eg
y
h
a
s
t
h
e
b
es
t
ev
al
u
at
i
o
n
res
u
l
t
s
i
n
c
l
as
s
e
x
t
r
a
v
er
s
i
o
n
/
i
n
t
ro
v
er
s
i
o
n
(E
/
I)
o
f
0
.
5
6
2
an
d
cl
a
s
s
feel
i
n
g
/
t
h
i
n
k
i
n
g
(F/
T
)
o
f
0
.
5
3
8
.
t
h
en
,
t
h
e
BE
RT
+
L
ST
M
ap
p
ro
ac
h
s
t
rat
e
g
y
h
as
t
h
e
h
i
g
h
e
s
t
acc
u
racy
f
o
r
t
h
e
i
n
t
u
i
t
i
on
/
s
e
n
s
i
n
g
(
N
/
S
)
cl
as
s
o
f
0
.
5
4
3
an
d
j
u
d
g
i
n
g
/
p
erce
i
v
i
n
g
(J
/
P)
o
f
0
.
5
3
2
.
K
e
y
w
o
r
ds
:
B
E
R
T
F
ine
tuni
ng
M
ye
r
s
-
B
r
iggs
type
indi
c
a
tor
P
e
r
s
ona
li
ty
de
tec
ti
on
S
e
que
nc
e
lea
r
ning
T
witt
e
r
(
X)
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
e
tno
Kus
umaningr
um
De
pa
r
tm
e
nt
of
I
nf
or
mat
ics
,
F
a
c
ult
y
o
f
S
c
ienc
e
a
nd
M
a
thema
ti
c
s
,
Unive
r
s
it
a
s
Dipone
gor
o
S
t.
P
r
of
.
J
a
c
ub
R
a
is
,
Unive
r
s
it
a
s
Dipone
gor
o,
T
e
m
ba
lang,
S
e
mar
a
ng
50275
,
I
ndone
s
ia
E
mail
:
r
e
tno@l
ive.
undip
.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
P
e
r
s
ona
li
ty
r
e
f
e
r
s
to
a
n
ind
ivi
dua
l's
typi
c
a
l
be
ha
vior
a
l,
e
mot
ional
,
a
nd
c
ognit
ive
pa
tt
e
r
ns
that
a
r
e
mos
tl
y
dis
playe
d
whe
n
e
nga
ging
with
other
s
.
Ana
lyzing
pe
r
s
ona
li
ty
ha
s
a
tt
r
a
c
ted
a
lot
of
int
e
r
e
s
t.
T
his
ha
s
r
e
s
ult
e
d
in
va
r
ious
a
r
e
a
s
s
uc
h
a
s
r
e
c
r
uit
ment,
a
dve
r
ti
s
ing,
a
nd
ma
r
ke
ti
ng.
T
he
s
e
f
ields
s
tudy
how
pe
r
s
ona
li
ty
inf
luenc
e
s
dif
f
e
r
e
nt
a
s
pe
c
ts
to
e
nha
nc
e
the
e
f
f
e
c
ti
ve
ne
s
s
of
s
tr
a
tegie
s
.
Unde
r
s
tanding
a
n
ind
ivi
dua
l's
pe
r
s
ona
li
ty
pr
ovides
ins
ight
s
int
o
their
ge
ne
r
a
l
c
ha
r
a
c
ter
is
ti
c
s
or
a
tt
it
ude
s
,
ther
e
by
maximi
z
ing
their
potential
[
1]
.
T
his
knowle
dge
a
ls
o
holds
s
igni
f
ica
nc
e
in
the
c
a
r
e
e
r
f
ield.
R
e
s
e
a
r
c
he
r
s
in
[
2]
,
[
3]
s
ugge
s
ts
that
indi
viduals
pe
r
f
o
r
m
be
t
ter
whe
n
their
pe
r
s
ona
li
ty
a
li
gns
with
their
job,
making
it
e
a
s
ier
to
a
da
pt
to
t
he
wor
k
e
nvir
onment
without
r
e
qui
r
ing
e
xtens
ive
phys
ica
l
a
bil
it
ies
.
He
r
r
e
t
al.
[
4]
s
uppor
ts
the
idea
that
in
divi
dua
ls
wor
k
be
s
t
in
r
oles
with
lowe
r
phys
ica
l
de
mands
.
T
he
method
to
dis
c
ove
r
pe
r
s
ona
li
ty
is
by
c
ompl
e
ti
ng
the
que
s
ti
onna
ir
e
f
o
r
the
pe
r
s
ona
li
ty
tes
t.
A
wide
ly
us
e
d
tool
f
or
a
s
s
e
s
s
ing
pe
r
s
ona
li
ty
is
M
y
e
r
s
-
B
r
iggs
type
indi
c
a
tor
(
M
B
T
I
)
.
M
B
T
I
is
f
ounde
d
on
C
a
r
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
339
5
-
3403
3396
J
ung's
pe
r
s
ona
l
it
y
theo
r
y
[
5]
a
nd
e
m
ploys
f
ou
r
-
f
a
c
tor
mode
l
inc
ludi
ng:
e
xt
r
a
ve
r
s
ion
(
E
)
o
r
i
ntr
ove
r
s
ion
(
I
)
,
int
u
it
i
on
(
N
)
or
s
e
ns
in
g
(
S
)
,
f
e
e
l
ing
(
F
)
o
r
th
inki
ng
(
T
)
,
a
nd
jud
ging
(
J
)
o
r
pe
r
c
e
ivi
n
g
(
P
)
.
H
owe
ve
r
,
tr
a
dit
i
ona
l
pe
r
s
ona
li
t
y
tes
ts
ha
ve
li
m
it
a
t
ions
a
s
r
e
s
p
ons
e
s
m
a
y
n
ot
a
lw
a
ys
be
c
ons
is
tent
d
ue
to
r
a
ndo
m
o
r
h
a
pha
z
a
r
d
a
ns
we
r
i
ng
,
le
a
din
g
to
va
r
iable
p
r
e
di
c
ted
r
e
s
u
lt
s
.
As
a
r
e
s
ul
t,
t
his
s
t
udy
s
ugge
s
ts
us
i
ng
s
oc
ial
med
ia
,
pa
r
ti
c
ula
r
l
y
X
(
c
omm
onl
y
known
a
s
T
wit
ter
)
,
to
de
tec
t
pe
r
s
ona
li
ty
t
r
a
it
s
.
X
pr
ovides
a
pla
tf
or
m
f
o
r
us
e
r
s
to
in
ter
a
c
t
,
e
xpr
e
s
s
tho
ugh
ts
a
n
d
f
e
e
l
ings
th
r
o
ugh
twe
e
ts
,
wh
ic
h
i
ndi
r
e
c
t
ly
r
e
ve
a
l
a
s
pe
c
ts
o
f
the
ir
pe
r
s
ona
l
it
y
[
6
]
.
S
e
ve
r
a
l
s
tudi
e
s
ha
ve
e
xa
mi
ne
d
the
us
e
of
X
da
ta
f
or
M
B
T
I
pe
r
s
ona
li
ty
de
tec
ti
on
us
ing
na
tur
a
l
langua
ge
pr
oc
e
s
s
ing
(
NL
P
)
.
T
he
a
ppr
oa
c
he
s
f
o
r
M
B
T
I
pe
r
s
ona
li
ty
c
las
s
if
ica
ti
on
mos
tl
y
invol
ve
binar
y
c
las
s
if
ica
ti
on.
F
r
kovic
e
t
al
.
[
7]
de
mons
tr
a
ted
tha
t
the
b
inar
y
c
las
s
if
ica
ti
on
a
ppr
oa
c
h
is
mo
r
e
e
f
f
e
c
ti
ve
than
the
mul
ti
c
las
s
c
las
s
if
ica
ti
on
a
ppr
oa
c
h.
T
he
r
e
s
e
a
r
c
he
r
s
ini
ti
a
ll
y
us
e
d
s
yntac
ti
c
a
na
lys
is
f
e
a
tur
e
s
a
nd
n
-
gr
a
m
c
ha
r
a
c
ter
is
ti
c
s
with
c
las
s
ica
l
mac
hine
lea
r
ning
methods
.
T
he
r
e
s
e
a
r
c
h
in
thi
s
f
ield
a
ls
o
uti
li
z
e
d
c
las
s
ica
l
mac
hine
lea
r
ning
methods
[
8]
,
[
9]
,
[
10
]
in
M
B
T
I
pe
r
s
ona
li
ty
de
tec
ti
on
.
How
e
ve
r
,
c
las
s
ica
l
mac
hine
lea
r
ning
methods
c
ha
ll
e
nge
s
in
f
e
a
tur
e
e
xtr
a
c
ti
on,
a
s
mi
s
s
ing
or
incomplete
f
e
a
tur
e
s
c
a
n
lea
d
to
s
ubopti
mal
output
s
[
11]
.
T
his
l
im
it
a
ti
on
c
a
n
be
ove
r
c
ome
by
us
ing
de
e
p
lea
r
ning
methods
,
pa
r
ti
c
ular
ly
s
e
que
nti
a
l
-
ba
s
e
d
a
r
c
hit
e
c
tur
e
s
li
ke
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
(
R
NN
s
)
.
S
ince
R
NN
s
c
a
n
a
utom
a
ti
c
a
ll
y
e
xtr
a
c
t
f
e
a
t
ur
e
s
a
nd
c
ons
ider
s
e
mantic
de
pe
nde
nc
ies
,
making
it
s
upe
r
ior
to
c
las
s
ica
l
mac
hine
lea
r
ning
methods
.
T
his
ha
s
be
e
n
pr
ove
n
in
r
e
s
e
a
r
c
h
by
[
12]
,
[
13]
.
F
oll
owing
the
a
dva
nc
e
ments
in
de
e
p
lea
r
ning
,
the
bidi
r
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
ntations
f
r
om
tr
a
ns
f
or
mer
s
(
B
E
R
T
)
is
a
notew
or
thy
br
e
a
kthr
ou
gh
in
NL
P
.
B
E
R
T
e
mpl
oys
the
tr
a
ns
f
or
mer
s
f
r
a
mew
or
k,
s
im
il
a
r
ly
to
the
ge
ne
r
a
ti
ve
pr
e
-
tr
a
ined
t
r
a
ns
f
or
mer
(
GP
T
)
a
nd
lar
ge
langua
ge
model
meta
-
AI
(
L
L
a
M
A)
.
T
he
a
r
c
hit
e
c
tur
e
r
e
li
e
s
on
s
e
lf
-
a
tt
e
nti
on
mec
ha
nis
m,
e
na
bli
ng
e
ve
r
y
token
to
be
e
va
luate
d
a
longs
ide
a
ll
other
tokens
a
t
the
s
a
me
ti
me
[
14]
.
As
a
r
e
s
ult
,
a
tt
e
nti
on
we
ight
s
be
twe
e
n
tokens
a
r
e
c
a
l
c
ulate
d
,
e
na
bli
ng
the
model
to
a
c
c
e
s
s
inf
or
mation
f
r
om
a
ll
input
s
.
B
E
R
T
e
mpl
oys
the
e
nc
ode
r
s
tr
uc
tur
e
of
tr
a
ns
f
o
r
mer
s
,
c
a
ptur
in
g
c
ontext
f
r
om
both
d
ir
e
c
ti
ons
to
c
ompr
e
he
nd
text
tho
r
oug
hly,
making
it
s
uit
a
ble
f
or
c
las
s
if
ica
ti
on.
S
e
ve
r
a
l
pr
e
vious
s
tudi
e
s
ha
ve
e
mpl
o
ye
d
B
E
R
T
f
or
M
B
T
I
pe
r
s
ona
l
it
y
de
tec
ti
on.
F
o
r
ins
tanc
e
,
r
e
s
e
a
r
c
h
by
[
15]
,
[
16]
uti
li
z
e
d
B
E
R
T
a
s
a
wor
d
e
mbedding
a
ppr
oa
c
h.
Additi
ona
ll
y,
r
e
s
e
a
r
c
he
r
s
by
[
17]
,
[
18
]
c
onduc
ted
r
e
s
e
a
r
c
h
by
f
ine
-
tuni
ng
B
E
R
T
to
c
las
s
if
y
the
f
ou
r
di
f
f
e
r
e
nt
M
B
T
I
dim
e
ns
ions
.
I
n
c
ontr
a
s
t
to
e
mpl
oying
B
E
R
T
e
xc
lu
s
ively
a
s
a
wor
d
e
mbedding,
whic
h
de
pe
nd
on
ve
c
to
r
s
ge
ne
r
a
ted
f
r
om
pr
e
-
tr
a
ined
models
,
f
ine
-
tuni
ng
B
E
R
T
invol
ve
s
r
e
tr
a
ini
ng
o
r
t
r
a
ns
f
e
r
r
ing
the
knowle
dge
f
r
o
m
th
e
B
E
R
T
model
uti
l
izing
a
s
pe
c
if
ic
tas
k
da
tas
e
t
[
19]
,
a
nd
s
ubs
e
que
ntl
y
incor
por
a
ti
ng
a
f
ull
y
c
onne
c
ted
lay
e
r
that
maps
the
r
e
p
r
e
s
e
ntation
r
e
s
ult
s
int
o
the
i
ntende
d
output
.
How
e
ve
r
,
a
lt
hough
B
E
R
T
e
xc
e
ls
in
unde
r
s
tanding
c
ontext,
it
s
ti
ll
ha
s
li
mi
tations
in
c
a
ptur
ing
wor
d
or
de
r
a
nd
s
tyl
is
ti
c
va
r
iations
.
T
he
r
e
f
o
r
e
,
thi
s
s
tudy
pr
opos
e
d
int
e
gr
a
ti
ng
f
ine
-
tuni
ng
B
E
R
T
with
R
NN
s
to
e
xa
mi
ne
the
im
pa
c
t
of
e
nha
nc
e
d
c
ontext
modeling
.
T
he
pr
opos
e
d
method
a
im
s
to
im
pr
ove
model
pe
r
f
or
manc
e
by
unde
r
s
tanding
the
global
c
ontext
of
wor
ds
a
nd
c
a
ptur
ing
de
e
pe
r
mea
ning
r
e
late
d
t
o
M
B
T
I
pe
r
s
ona
li
ty.
T
he
r
e
mainde
r
of
the
pa
pe
r
is
or
g
a
nize
d
a
s
f
oll
ows
:
s
e
c
ti
on
2
outl
ines
the
methodology,
pr
o
vidi
ng
a
c
ompr
e
he
ns
ive
ove
r
view
of
the
da
ta,
f
i
ne
-
tuni
ng
B
E
R
T
f
or
pe
r
s
ona
li
ty
de
tec
ti
on,
a
nd
it
s
ugge
s
t
the
int
e
gr
a
ti
on
of
R
NN
s
.
S
e
c
ti
on
3
e
xplai
n
e
xpe
r
im
e
nt
s
e
tup
a
nd
a
dis
c
us
s
ion
of
thos
e
r
e
s
ult
s
,
th
is
s
tudy
im
pl
i
c
a
ti
on
a
nd
f
u
tur
e
r
e
s
e
a
r
c
h.
F
inally
,
s
e
c
ti
on
4
c
ont
a
ins
the
c
onc
lus
ion
of
thi
s
s
tudy.
2.
M
E
T
HO
DOL
OG
Y
T
he
methodology
s
tar
ts
with
c
oll
e
c
ti
ng
da
ta
by
s
c
r
a
ping
twe
e
ts
f
r
om
the
X
s
oc
ial
medi
a
platf
or
m
.
Onc
e
the
da
ta
is
c
oll
e
c
ted,
it
is
pr
e
pr
oc
e
s
s
e
d
be
f
or
e
be
ing
input
int
o
the
model.
T
he
pe
r
f
or
manc
e
of
the
model
is
e
va
luate
d
by
a
n
a
c
c
ur
a
c
y
metr
ic.
F
igur
e
1
de
picts
the
pha
s
e
s
of
the
pr
opos
e
d
s
tudy,
a
nd
th
e
de
tails
of
e
a
c
h
pha
s
e
will
be
e
xplaine
d
in
the
s
ubs
e
que
nt
s
ub
-
s
e
c
ti
on
s
.
2.
1.
Dat
as
e
t
T
he
da
tas
e
t
c
ontains
twe
e
t
da
ta
f
r
om
I
ndone
s
ian
X
us
e
r
s
who
ha
ve
s
ha
r
e
d
their
r
e
s
ult
s
of
pe
r
s
ona
li
ty
tes
ts
f
r
om
va
r
ious
pe
r
s
ona
li
ty
tes
t
s
e
r
vice
s
16pe
r
s
ona
li
ti
e
s
.
c
om.
T
he
da
tas
e
t
c
ons
is
ts
of
twe
e
t
da
ta
f
r
om
I
ndone
s
ian
X
us
e
r
s
who
ha
ve
s
ha
r
e
d
thei
r
pe
r
s
ona
li
ty
tes
t
r
e
s
ult
s
f
r
om
the
we
bs
it
e
16pe
r
s
ona
li
ti
e
s
.
c
om.
T
he
50
mos
t
r
e
c
e
nt
twe
e
ts
f
r
om
thes
e
us
e
r
s
will
be
r
e
tr
ieve
d
us
ing
the
X
a
ppli
c
a
ti
on
pr
ogr
a
mm
ing
int
e
r
f
a
c
e
(
API
)
li
br
a
r
y.
At
the
s
a
me
ti
me
,
the
s
ha
r
e
d
pe
r
s
ona
li
ty
te
s
t
r
e
s
ult
s
will
be
us
e
d
a
s
da
ta
labe
ls
.
Af
ter
that
,
da
ta
labe
ls
will
be
divi
de
d
in
to
f
our
c
a
tegor
ies
:
E
/I
,
N/S
,
F
/T
,
a
nd
J
/P
.
5120
da
ta
we
r
e
s
uc
c
e
s
s
f
ull
y
c
oll
e
c
t
e
d
with
ba
lanc
e
d
c
las
s
e
s
f
or
a
ll
M
B
T
I
pe
r
s
ona
li
ty
types
.
F
igur
e
2
s
hown
the
lengt
h
or
the
da
ta.
I
t
c
a
n
be
s
e
e
n
that
s
ome
us
e
r
s
ha
v
e
a
token
length
of
les
s
than
200,
whic
h
mea
ns
that
thos
e
us
e
r
s
ha
ve
les
s
than
or
e
qua
l
to
4
tokens
in
e
a
c
h
of
their
twe
e
ts
.
On
t
he
other
ha
nd,
s
ome
us
e
r
s
ha
ve
long
twe
e
ts
with
mor
e
tha
n
1,
000
tokens
,
s
o
the
s
e
us
e
r
s
ha
ve
mor
e
than
20
tokens
or
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
F
ine
-
tuni
ng
b
idi
r
e
c
ti
onal
e
nc
ode
r
r
e
pr
e
s
e
ntat
ions
fr
om
tr
ans
for
me
r
s
for
…
(
Se
lvi
F
it
r
ia
K
hoe
r
unnis
a)
3397
wor
ds
pe
r
twe
e
t.
M
e
a
nwhile,
mos
t
da
ta
dis
tr
ibut
i
on
is
a
r
ound
the
300
r
a
nge
.
T
hus
,
on
a
ve
r
a
ge
,
o
ne
twe
e
t
c
ontains
5
to
7
wo
r
ds
.
F
r
om
F
igur
e
2
,
i
t
c
a
n
be
s
e
e
n
that
the
da
ta
is
long
a
nd
quit
e
c
ompl
e
x
.
F
igur
e
1.
R
e
s
e
a
r
c
h
methodology
F
igur
e
2.
Dis
tr
ibut
ion
of
length
token
da
ta
2.
2.
P
r
e
p
r
oc
e
s
s
in
g
T
he
X
da
ta
r
e
s
ult
ing
f
r
om
c
r
a
wling
is
da
ta
that
h
a
s
much
nois
e
a
nd
is
uns
tr
uc
tur
e
d
,
s
o
it
r
e
qui
r
e
s
pr
e
pr
oc
e
s
s
ing
to
r
e
duc
e
the
dim
e
ns
ions
of
the
d
a
ta
dur
ing
t
r
a
ini
ng
[
20]
,
[
21
]
.
All
the
pr
e
pr
oc
e
s
s
ing
s
teps
c
a
r
r
ied
out
us
ing
the
I
ndoNL
P
li
br
a
r
y
.
P
r
e
pr
oc
e
s
s
ing
is
e
s
s
e
nti
a
l
be
c
a
us
e
it
a
ll
ows
the
model
identi
f
y
dis
ti
nc
t
pa
tt
e
r
ns
in
the
da
ta,
e
na
bli
ng
the
a
na
lys
is
a
nd
c
las
s
if
ica
ti
on
of
pe
r
s
ona
li
ti
e
s
.
T
he
pr
e
pr
oc
e
s
s
i
ng
s
teps
c
onduc
ted
in
or
de
r
a
r
e
a
s
f
ol
lows
:
i)
R
e
move
X
s
pe
c
ial
c
ha
r
a
c
ter
s
,
s
uc
h
a
s
ment
ions
,
ha
s
htags
,
a
n
d
UR
L
s
,
be
c
a
us
e
the
y
a
r
e
mea
n
ing
les
s
[
2
2]
.
ii)
C
a
s
e
f
oldi
ng,
c
onve
r
ted
f
r
om
a
ll
letter
s
to
lowe
r
c
a
s
e
.
Additi
ona
ll
y,
a
ll
letter
s
a
r
e
c
onve
r
ted
to
lowe
r
c
a
s
e
to
e
ns
ur
e
c
ons
is
tenc
y
a
nd
pr
e
ve
nt
s
igni
f
ica
nt
va
r
ia
ti
ons
in
wor
d
ve
c
tor
s
.
iii)
C
onve
r
t
e
moj
i
or
e
mot
icon
to
s
tr
ings
.
iv)
Da
ta
c
lea
ning,
whic
h
include
s
r
e
move
wor
d
e
lon
ga
ti
on,
s
lang
wor
ds
,
a
nd
s
top
wor
ds
.
S
top
wo
r
ds
a
r
e
f
r
e
que
ntl
y
mea
ningl
e
s
s
,
s
o
the
model
c
a
n
f
oc
us
on
ly
on
e
s
s
e
nti
a
l
wor
ds
that
c
ont
r
ibut
e
mor
e
to
the
te
xt's
mea
ning
[
23]
.
2.
3.
F
in
e
t
u
n
in
g
B
E
RT
B
E
R
T
is
a
ve
r
s
a
ti
le
model
that
t
r
a
ins
bidi
r
e
c
ti
ona
l
r
e
pr
e
s
e
ntations
o
f
unlabe
led
text
[
14
]
.
B
E
R
T
ha
s
gr
e
a
tl
y
im
pr
ove
d
NL
P
by
e
f
f
e
c
ti
ve
ly
unde
r
s
tanding
c
ontext
a
nd
s
e
mantics
in
text,
a
na
lyzing
inf
or
mation
f
r
om
both
dir
e
c
ti
ons
[
24]
,
[
25
]
.
B
E
R
T
c
a
n
be
im
pleme
nted
us
ing
two
a
ppr
oa
c
he
s
:
f
e
a
tur
e
-
ba
s
e
d
a
nd
f
ine
-
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
339
5
-
3403
3398
tuni
ng.
T
he
f
e
a
tur
e
-
ba
s
e
d
model
is
p
r
e
s
e
r
ve
d,
a
nd
the
outpu
t
is
a
f
e
a
tur
e
ve
c
tor
f
or
the
s
ubs
e
que
nt
c
las
s
if
ica
ti
on
model
[
26
]
,
thi
s
pr
oc
e
s
s
a
ls
o
known
a
s
wor
d
e
mbedding
pr
oc
e
s
s
.
T
he
r
e
s
ult
ing
ve
c
tor
will
be
s
e
nt
thr
ough
the
s
pe
c
if
ied
c
las
s
if
ier
.
I
n
c
ont
r
a
s
t,
f
ine
-
tuni
ng
r
e
tr
a
ins
the
model
to
s
olve
a
mor
e
s
pe
c
if
ic
pr
oblem
by
mod
if
ying
or
a
djus
ti
ng
the
model
a
r
c
hi
tec
tur
e
,
s
howc
a
s
ing
B
E
R
T
's
a
da
ptabili
ty
to
dif
f
e
r
e
nt
tas
ks
.
As
e
x
plai
ne
d
be
f
o
r
e
,
t
his
s
tu
dy
i
mpl
e
mente
d
f
ine
t
uning
B
E
R
T
.
F
i
r
s
t
a
t
a
l
l
,
the
da
tas
e
t
r
e
qui
r
e
d
to
f
it
the
B
E
R
T
i
nput
f
o
r
ma
t,
ne
c
e
s
s
it
a
ti
n
g
a
to
ke
niza
ti
on
pr
oc
e
s
s
to
a
l
ign
wit
h
the
pr
e
-
t
r
a
ine
d
mo
de
l
.
I
t
invol
ve
d
a
dding
u
niqu
e
tokens
t
o
e
a
c
h
s
e
ntenc
e
a
n
d
c
onve
r
ti
ng
t
he
da
t
a
in
to
ve
c
t
or
s
.
F
i
ne
-
tuni
ng
wa
s
c
r
uc
ial
to
a
djus
t
a
ll
pa
r
a
mete
r
s
pr
e
c
is
e
ly
.
S
p
e
c
ial
s
y
mbol
s
s
uc
h
a
s
[
C
L
S
]
a
n
d
[
S
E
P
]
we
r
e
a
dde
d
a
t
the
be
g
inn
ing
a
n
d
e
n
d
of
e
a
c
h
inp
ut
,
wi
th
p
a
ddi
ng
us
e
d
to
e
ns
ur
e
uni
f
o
r
m
da
ta
le
ngt
h
[
14]
.
T
he
[
C
L
S
]
t
oke
n
wa
s
incl
ude
d
in
the
downs
t
r
e
a
m
tas
k
a
s
a
n
a
gg
r
e
ga
te
r
e
p
r
e
s
e
nta
ti
o
n
s
u
mm
a
r
izin
g
the
inpu
t
s
e
q
ue
nc
e
in
f
o
r
mat
ion
.
T
o
f
ine
-
tune
t
he
ve
c
to
r
c
las
s
i
f
ica
ti
o
n
m
ode
l
r
e
late
d
to
[
C
L
S
]
,
it
wa
s
input
in
to
the
e
nc
ode
r
be
f
or
e
a
ddi
ng
a
ne
u
r
a
l
ne
t
wo
r
k
laye
r
a
bove
t
he
ou
tpu
t
lay
e
r
.
F
igu
r
e
3
s
hows
t
he
ou
t
put
la
ye
r
c
a
n
be
in
teg
r
a
ted
wit
h
ot
he
r
a
r
c
h
it
e
c
t
ur
e
s
.
As
il
lus
t
r
a
t
e
d
in
F
ig
ur
e
s
3
(
a
)
to
3
(
d
)
,
the
i
nteg
r
a
t
ion
of
f
ine
-
tu
ning
B
E
R
T
wit
h
the
R
NN
s
u
ti
l
ize
d
i
n
th
is
s
tu
dy
.
T
he
s
e
l
e
c
ted
R
NN
s
inc
lude
va
ni
ll
a
R
N
N,
lon
g
s
h
or
t
-
te
r
m
memo
r
y
(
L
S
T
M
)
,
a
nd
ga
ted
r
e
c
u
r
r
e
nt
uni
t
(
GR
U
)
.
Apa
r
t
f
r
om
tha
t
,
e
xpe
r
im
e
nts
wil
l
a
ls
o
be
c
a
r
r
ied
o
ut
wit
h
the
ba
s
e
laye
r
o
f
B
E
R
T
,
w
it
h
o
nly
a
f
ull
y
c
onne
c
te
d
laye
r
a
d
de
d
a
s
a
c
o
mpa
r
is
on
.
A
de
s
c
r
ipt
ion
o
f
e
a
c
h
R
NN
me
thod
is
e
x
plai
ne
d
in
t
he
f
o
ll
ow
ing
s
ubs
e
c
ti
on.
(
a
)
(
b)
(
c
)
(
d)
F
igur
e
3.
Ar
c
hit
e
c
tur
e
of
f
ine
tuni
ng
(
a
)
B
E
R
T
ba
s
e
f
ull
y
c
onne
c
ted
laye
r
,
(
b)
B
E
R
T
+
va
nil
la
R
NN
,
(
c
)
B
E
R
T
+
L
S
T
M
,
a
nd
(
d
)
B
E
R
T
+
GR
U
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
F
ine
-
tuni
ng
b
idi
r
e
c
ti
onal
e
nc
ode
r
r
e
pr
e
s
e
ntat
ions
fr
om
tr
ans
for
me
r
s
for
…
(
Se
lvi
F
it
r
ia
K
hoe
r
unnis
a)
3399
T
he
lowe
r
laye
r
s
of
B
E
R
T
c
ontain
mor
e
ge
ne
r
a
li
z
e
d
inf
or
mation
,
whe
r
e
a
s
the
uppe
r
laye
r
s
f
oc
us
on
s
pe
c
if
ic
tas
ks
inf
or
mation.
I
n
f
ine
-
tuni
ng,
the
c
las
s
if
ica
ti
on
tas
k’
s
model
is
ini
ti
a
li
z
e
d
with
pr
e
-
tr
a
ined
pa
r
a
mete
r
s
,
a
nd
then
modi
f
ied
to
f
it
the
labe
led
da
ta;
s
e
ve
r
a
l
of
the
a
djus
tm
e
nts
mentioned
in
d
e
tail
a
r
e
e
xplaine
d
a
s
f
oll
ows
.
As
de
picte
d
in
F
igur
e
s
3
(
a
)
to
3(
d)
,
the
model
a
r
c
hit
e
c
tur
e
a
ppr
oa
c
h
is
B
E
R
T
-
ba
s
e
d
f
ull
y
c
onne
c
ted
laye
r
,
thi
s
a
r
c
hit
e
c
tur
e
us
e
s
a
r
e
pr
e
s
e
ntation
of
X
wor
ds
pr
ovided
by
B
E
R
T
(
las
t
hidden
laye
r
)
,
whic
h
is
dir
e
c
tl
y
c
onne
c
ted
to
the
f
ull
y
c
onne
c
ted
laye
r
without
a
ny
hidden
laye
r
s
.
T
he
a
c
ti
va
ti
on
f
unc
ti
on
is
a
ppli
e
d
to
the
f
inal
laye
r
f
or
c
las
s
if
ica
ti
on.
B
E
R
T
+
(
va
nil
la
R
NN
,
L
S
T
M
,
or
GR
U)
la
ye
r
.
T
he
a
r
c
hit
e
c
t
ur
e
us
e
s
wor
d
r
e
pr
e
s
e
ntations
ge
ne
r
a
ted
f
r
om
the
las
t
h
idden
laye
r
o
f
B
E
R
T
,
whic
h
is
c
onn
e
c
ted
to
the
R
NN
laye
r
.
I
t
pe
r
f
o
r
ms
R
NN
s
be
f
or
e
c
onne
c
ti
ng
int
o
the
f
u
ll
y
c
onne
c
ted
laye
r
a
nd
output
with
a
n
a
c
ti
va
ti
on
f
unc
ti
on.
T
he
a
r
c
hit
e
c
tur
e
of
thi
s
a
ppr
oa
c
h
a
ls
o
us
e
s
dr
opout
a
s
a
r
e
gulation
tec
hnique.
2.
4.
Vani
ll
a
r
e
c
c
u
r
e
n
t
n
e
u
r
al
n
e
t
wor
k
R
NN
s
is
a
type
o
f
ne
ur
a
l
ne
two
r
k
with
loops
.
R
NN
s
ha
s
memor
y
a
nd
a
ll
ows
it
to
s
tor
e
e
xis
ti
ng
inf
or
mation
[
27]
.
Va
nil
la
R
NN
is
a
type
of
R
NN
s
with
only
one
it
e
r
a
ti
on
,
mea
nin
g
that
va
nil
la
R
NN
c
a
n
only
s
tor
e
inf
or
mation
f
r
o
m
one
p
r
e
vious
s
tate
.
2.
5.
L
on
g
s
h
or
t
-
t
e
r
m
m
e
m
or
y
L
S
T
M
is
a
f
o
r
m
o
f
R
NN
s
that
a
ddr
e
s
s
e
s
the
li
m
it
a
ti
on
of
va
nil
la
R
NN
.
I
f
va
nil
la
R
NN
c
a
n
only
s
tor
e
inf
o
r
mation
f
r
om
one
pr
e
vious
s
tate
,
L
S
T
M
c
a
n
s
tor
e
in
f
or
mation
f
r
om
a
ll
pr
e
vious
s
t
a
tes
a
nd
ove
r
c
ome
long
-
ter
m
text
de
pe
nde
nc
y
[
28
]
,
[
29
]
.
F
e
a
tur
e
s
of
L
S
T
M
c
ons
is
t
of
memor
y
c
e
ll
s
a
nd
t
hr
e
e
ga
te
unit
s
(
input
ga
te
,
f
o
r
ge
t
ga
te,
a
nd
output
ga
te)
to
r
e
a
d,
s
tor
e
,
a
nd
upda
te
in
f
or
mation
.
2.
6.
Gat
e
d
r
e
c
u
r
r
e
n
t
u
n
it
GR
U
is
a
ls
o
a
n
im
pr
ove
d
a
r
c
hit
e
c
tur
e
ove
r
va
nil
la
R
NN
a
nd
c
a
n
ha
ndle
long
-
ter
m
de
pe
nde
nc
ie
s
of
text.
T
he
dis
ti
nc
ti
on
be
twe
e
n
GR
U
a
nd
L
S
T
M
is
f
ound
in
the
type
of
ga
te
they
pos
s
e
s
s
.
I
f
L
S
T
M
ha
s
thr
e
e
ga
tes
,
GR
U
only
ha
s
upda
te
ga
tes
a
nd
r
e
s
e
t
g
a
te
s
[
30]
,
[
31
]
.
T
he
upda
te
ga
te
is
a
mer
g
ing
input
ga
te
a
nd
f
or
ge
t
ga
te,
while
the
r
e
s
e
t
ga
te
s
e
ts
the
va
lue
f
r
om
the
pr
e
vious
s
tate
to
c
onti
nue
to
the
ne
xt
s
tate
.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
3.
1.
I
m
p
lem
e
n
t
at
ion
T
he
im
pleme
ntation
of
thi
s
s
tudy
us
e
d
the
p
re
-
tr
a
ined
I
ndoB
E
R
T
we
e
t
-
ba
s
e
-
unc
a
s
e
d
[
32]
f
o
r
m
I
ndoL
E
M
,
a
pr
e
-
tr
a
ined
langua
ge
model
f
or
I
ndon
e
s
ian
langua
ge
whic
h
ha
ve
409
M
tokens
.
I
ndoB
E
R
T
we
e
t
ha
s
be
e
n
tr
a
ined
ba
s
e
d
on
B
E
R
T
-
ba
s
e
-
unc
a
s
e
d
by
uti
li
z
ing
12
a
tt
e
nti
on
he
a
ds
,
12
hidden
laye
r
s
,
f
e
e
d
-
f
or
wa
r
d
hidden
laye
r
s
,
a
nd
180
e
poc
hs
[
14]
.
Af
ter
p
r
oc
e
s
s
ing,
the
da
ta
is
tokeniz
e
d
us
ing
the
s
a
me
a
ppr
oa
c
h
a
s
the
pr
e
-
tr
a
ined
model.
T
oke
niza
ti
on
not
only
s
e
pa
r
a
tes
punc
tuation
a
nd
r
e
moves
invalid
c
ha
r
a
c
ter
s
but
a
ls
o
pr
e
pa
r
e
s
the
da
ta
f
o
r
a
na
lys
is
.
T
he
uppe
r
li
m
it
f
or
s
e
ntenc
e
length
is
de
ter
mi
ne
d
to
be
512
tokens
a
c
c
or
ding
to
the
dis
tr
ibut
ion
of
token
lengths
in
the
da
tas
e
t.
I
f
a
n
input
is
s
hor
ter
than
thi
s
length,
z
e
r
os
a
r
e
a
dde
d
to
pa
d
it
;
i
f
it
e
xc
e
e
ds
thi
s
li
mi
t
,
it
is
tr
unc
a
ted
to
f
it
.
T
he
tokeniz
e
d
da
tas
e
t
is
s
e
pa
r
a
ted
int
o
thr
e
e
s
e
gments
:
tr
a
ini
ng
da
ta,
tes
ti
ng
da
ta,
a
nd
va
li
da
ti
on
da
ta
,
f
ol
lowing
a
n
80:20
divi
s
ion
f
or
t
r
a
ini
ng
a
nd
tes
ti
ng.
T
he
v
a
li
da
ti
on
da
ta
c
ons
is
t
s
of
10%
of
the
tr
a
ini
ng
da
ta
.
T
he
m
ode
l
wa
s
tr
a
ined
with
hype
r
pa
r
a
mete
r
s
:
ba
tch
s
iz
e
16
a
nd
e
poc
h
25.
T
he
d
r
opout
pr
oba
bil
it
y
wa
s
s
e
t
f
or
a
ll
l
a
ye
r
s
a
t
0.
5.
T
he
Ada
mW
opti
mi
z
e
r
uti
li
z
e
s
a
lea
r
ning
r
a
te
of
1e
-
5.
F
or
e
va
luation,
the
s
tudy
e
mpl
oys
a
c
onf
u
s
ion
matr
ix
a
long
with
a
c
c
ur
a
c
y
metr
ics
s
ince
the
da
tas
e
t
is
ba
lanc
e
d.
S
ubs
e
que
ntl
y,
a
ll
e
xpe
r
im
e
nts
we
r
e
c
onduc
ted
us
ing
the
T
4
GPU,
a
T
ur
ing
a
r
c
hit
e
c
tur
e
GPU
int
e
nde
d
to
e
nha
nc
e
the
in
f
e
r
e
nc
e
pr
oc
e
s
s
of
de
e
p
l
e
a
r
ning
models
.
3.
2.
Re
s
u
lt
a
n
d
d
is
c
u
s
s
ion
T
his
s
tudy
us
e
d
a
binar
y
a
ppr
oa
c
h
to
r
e
c
ognize
M
B
T
I
pe
r
s
ona
li
ty,
with
f
our
c
a
tegor
ies
c
ons
is
t
of
E
/I
,
N/
S
,
F
/
T
,
a
nd
J
/P
.
T
he
method
us
ing
f
ine
-
tuni
ng
B
E
R
T
int
e
g
r
a
ted
with
R
NN
s
.
T
a
ble
1
a
nd
F
igur
e
4
c
ompar
e
the
a
ve
r
a
ge
of
the
r
e
s
ult
s
.
F
igur
e
4
de
mo
ns
tr
a
tes
that
f
ine
-
tuni
ng
B
E
R
T
with
a
f
ull
y
c
onne
c
ted
laye
r
a
c
hieve
s
the
highes
t
a
ve
r
a
ge
a
c
c
ur
a
c
y
with
a
va
lue
of
0
.
533.
T
he
n
,
it
wa
s
f
oll
owe
d
by
B
E
R
T
+
va
nil
la
R
NN
of
0.
523
a
nd
B
E
R
T
+
L
S
T
M
with
a
va
lue
of
0
.
518.
M
or
e
ove
r
,
the
las
t
is
B
E
R
T
+
GR
U,
with
a
va
lue
of
0.
504
.
F
r
om
F
igu
r
e
4,
s
hown
that
a
ddit
ion
of
R
NN
s
a
r
c
hit
e
c
tur
e
a
f
ter
B
E
R
T
f
ine
-
tuni
ng
a
f
f
e
c
ts
the
mod
e
l
pr
e
diction
r
e
s
ult
s
.
B
E
R
T
with
a
t
r
a
ns
f
or
mer
s
ba
s
e
is
us
e
d
to
unde
r
s
tand
the
r
e
lations
hip
be
twe
e
n
wor
ds
in
the
text,
s
o
a
dding
R
NN
s
that
a
im
to
c
a
ptur
e
s
e
que
nc
e
s
will
li
ke
ly
dim
ini
s
h
the
incr
e
menta
l
va
lue
that
L
S
T
M
mi
ght
c
ontr
ibut
e
.
T
he
pos
s
ibi
li
ty
of
a
m
b
iguous
a
nd
ove
r
ly
s
hor
t
twe
e
ts
a
ls
o
make
s
mor
e
dif
f
icult
.
As
a
r
e
s
ult
,
int
e
gr
a
ti
on
with
R
NN
s
o
f
ten
f
a
il
s
to
im
pr
ove
pe
r
f
o
r
manc
e
.
M
e
a
nwhile,
with
f
ul
ly
c
onne
c
ted
laye
r
,
th
e
B
E
R
T
r
e
pr
e
s
e
ntation
r
e
s
ult
s
a
r
e
dir
e
c
tl
y
e
nter
e
d
int
o
a
s
im
ple
matr
ix
a
nd
mappe
d
to
the
de
s
ir
e
d
labe
l.
A
lt
hough,
a
s
s
e
e
n
in
T
a
ble
1
B
E
R
T
+
L
S
T
M
f
ine
-
tuni
ng
ha
s
the
be
s
t
a
c
c
ur
a
c
y
f
or
the
N/S
c
las
s
a
nd
the
J
/
P
c
las
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
339
5
-
3403
3400
with
va
lues
of
0.
543
a
nd
0.
532
.
F
or
the
other
two
c
las
s
e
s
,
the
be
s
t
a
c
c
ur
a
c
y
is
us
ing
the
B
E
R
T
b
a
s
e
f
ull
y
c
onne
c
ted
laye
r
me
thod,
wi
th
va
lues
of
0.
562
a
nd
0.
538.
T
a
ble
1.
E
xpe
r
i
ment
r
e
s
ult
F
in
e
t
uni
ng s
tr
a
te
gi
e
s
L
a
be
l
E
/I
N
/S
F
/T
J
/P
B
E
R
T
B
A
S
E
F
U
L
L
Y
C
O
N
N
E
C
T
E
D
L
A
Y
E
R
0.562
0.524
0.538
0.510
B
E
R
T
+
va
ni
ll
a
R
N
N
0.545
0.521
0.523
0.505
B
E
R
T
+
L
S
T
M
0.518
0.543
0.480
0.532
B
E
R
T
+
G
R
U
0.516
0.516
0.480
0.505
F
igur
e
4.
Ave
r
a
ge
of
e
xpe
r
im
e
nt
r
e
s
ult
T
he
E
/I
dim
e
ns
ion
r
e
f
e
r
s
to
how
a
pe
r
s
on
di
r
e
c
ts
e
ne
r
gy
a
nd
pa
ys
a
tt
e
nti
on.
T
his
c
a
n
be
obs
e
r
ve
d
thr
ough
their
int
e
r
a
c
ti
on
s
tyl
e
a
nd
e
moj
is
in
twe
e
ts
.
T
he
F
/T
dim
e
ns
ion
invol
ve
s
a
pr
e
f
e
r
e
nc
e
f
or
de
c
is
ion
-
making
ba
s
e
d
on
objec
ti
ve
pr
inciples
r
a
ther
than
p
e
r
s
ona
l
f
e
e
li
ngs
.
W
he
n
a
na
lyzing
twe
e
ts
,
thi
s
dim
e
ns
ion
is
r
e
f
lec
ted
in
the
c
hoice
of
wor
ds
a
nd
tone,
whe
t
he
r
the
a
uthor
us
e
s
f
a
c
tual
langua
ge
or
opts
f
or
a
mor
e
e
mpathe
ti
c
a
nd
e
mot
ional
tone
.
S
o,
E
/I
a
nd
F
/
T
do
not
r
e
quir
e
the
a
ddit
ion
of
L
S
T
M
be
c
a
us
e
they
a
r
e
s
uf
f
icie
nt
to
c
a
ptur
e
f
r
e
que
ntl
y
a
ppe
a
r
ing
wor
ds
a
nd
unde
r
s
tand
r
e
lations
hips
dir
e
c
tl
y
without
the
ne
e
d
f
or
a
ti
me
s
e
que
nc
e
.
T
he
r
e
f
or
e
,
a
f
ull
y
c
onne
c
ted
l
a
ye
r
is
s
uf
f
icie
nt.
I
n
c
ontr
a
s
t,
the
N/S
dim
e
ns
ion
f
oc
u
s
e
s
on
a
pe
r
s
on's
pr
e
f
e
r
e
nc
e
f
or
a
c
quir
ing
inf
or
mation
via
the
f
ive
s
e
ns
e
s
or
thr
ough
pa
tt
e
r
ns
a
nd
pos
s
ibi
l
it
ies
.
I
n
twe
e
ts
,
thi
s
dim
e
ns
ion
pe
r
tains
to
how
inf
or
mati
on
is
pr
oc
e
s
s
e
d,
whe
ther
the
twe
e
ts
c
onve
y
r
e
a
li
ty
us
ing
s
tr
a
ight
f
or
wa
r
d
langua
ge
or
tend
to
be
mo
r
e
c
onc
e
ptual
a
nd
s
pe
c
ulative.
L
a
s
tl
y,
the
J
/P
dim
e
ns
ion
r
e
late
s
to
a
n
indi
vidual's
li
f
e
s
tyl
e
pr
e
f
e
r
e
nc
e
f
or
e
it
he
r
s
tr
uc
t
ur
e
a
nd
de
f
ini
tene
s
s
or
f
lexibil
it
y
a
nd
a
da
ptabili
ty
.
On
the
twe
e
t,
thi
s
dim
e
ns
ion
is
r
e
f
lec
ted
in
thei
r
c
omm
un
i
c
a
ti
on
s
tyl
e
,
whic
h
may
be
e
it
he
r
s
ys
tema
ti
c
a
nd
f
or
mal
o
r
s
pontane
ous
,
e
xplor
a
tor
y,
a
nd
of
ten
us
ing
a
bbr
e
vi
a
ti
ons
.
I
t
e
xplains
why
B
E
R
T
+
L
S
T
M
is
mor
e
e
f
f
e
c
ti
ve
f
or
the
N/S
a
nd
J
/P
dim
e
ns
ions
.
I
t
invol
ve
s
mor
e
c
o
mpl
e
x
s
e
ntenc
e
s
tr
uc
tur
e
s
a
nd
a
be
tt
e
r
unde
r
s
tand
ing
of
the
f
low
of
langua
ge
ove
r
t
im
e
.
C
ompar
e
d
with
pr
e
vious
s
tudi
e
s
Da
tt
a
e
t
al
.
[
33
]
wa
s
us
ing
B
E
R
T
f
oll
owe
d
by
r
a
ndom
f
o
r
e
s
t
(
R
F
)
a
nd
e
xtr
e
me
gr
a
dient
boos
ti
ng
(
XG
B
)
a
s
c
las
s
if
ier
s
f
or
M
B
T
I
pe
r
s
ona
li
ty
de
te
c
ti
on
on
X
da
ta.
T
hos
e
s
tudi
e
s
r
e
por
ted
the
be
s
t
a
c
c
ur
a
c
ies
of
0
.
441
a
nd
0.
424
f
or
e
a
c
h
c
las
s
if
ier
.
I
n
c
ontr
a
s
t,
the
pr
opos
e
d
meth
od
in
thi
s
s
tudy
a
c
hieve
d
a
be
tt
e
r
a
c
c
ur
a
c
y
of
0.
562
.
T
he
s
tatic
e
mbeddings
pr
oduc
e
d
by
the
B
E
R
T
model
in
thos
e
s
tudi
e
s
may
no
t
be
f
ull
y
opti
m
ize
d
f
or
c
e
r
tain
s
pe
c
if
ic
tas
ks
,
a
nd
tr
e
e
-
ba
s
e
d
c
la
s
s
if
ier
s
do
not
inher
e
ntl
y
r
e
c
ognize
s
e
que
nti
a
l
de
pe
nde
nc
ies
.
I
n
c
onc
lus
ion,
the
B
E
R
T
-
ba
s
e
d
f
ull
y
c
onne
c
ted
laye
r
c
o
ns
is
tently
outper
f
or
ms
the
o
ther
s
whe
n
it
c
omes
to
d
ir
e
c
tl
y
c
a
ptur
ing
the
mea
ning
of
text
by
identi
f
ying
f
r
e
que
ntl
y
oc
c
ur
r
ing
wor
ds
.
I
n
c
ontr
a
s
t,
the
B
E
R
T
+
L
S
T
M
e
xc
e
ls
in
unde
r
s
tanding
mo
r
e
c
ompl
e
x
s
e
ntenc
e
s
tr
uc
tur
e
s
a
nd
mana
ging
tempor
a
l
s
e
que
nc
e
s
.
T
he
r
e
f
or
e
,
t
he
s
e
lec
ti
on
of
the
be
s
t
model
c
a
n
be
made
b
y
s
im
ply
c
ons
ider
ing
th
e
highes
t
a
c
c
ur
a
c
y
f
o
r
e
a
c
h
labe
l
,
e
li
mi
na
ti
ng
the
ne
e
d
f
or
s
tatis
ti
c
a
l
tes
ti
ng.
A
lt
hough,
thi
s
s
tudy
inves
ti
ga
tes
the
im
pa
c
t
o
f
in
tegr
a
ti
ng
B
E
R
T
f
ine
-
tuni
ng
wi
th
R
NN
s
by
only
a
pplyi
ng
f
ull
y
c
onne
c
ted
laye
r
.
How
e
ve
r
,
f
ur
ther
c
ompr
e
he
ns
iv
e
s
tudi
e
s
a
r
e
ne
c
e
s
s
a
r
y
to
e
ns
ur
e
that
the
int
e
gr
a
ti
on
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
F
ine
-
tuni
ng
b
idi
r
e
c
ti
onal
e
nc
ode
r
r
e
pr
e
s
e
ntat
ions
fr
om
tr
ans
for
me
r
s
for
…
(
Se
lvi
F
it
r
ia
K
hoe
r
unnis
a)
3401
R
NN
s
ha
s
a
s
igni
f
ica
ntl
y
i
mpac
ts
on
e
va
luation
r
e
s
ult
s
,
pa
r
ti
c
ular
ly
with
r
e
ga
r
d
to
the
s
uit
a
bil
it
y
o
f
the
da
ta
type
or
s
tr
uc
tur
e
us
e
d
wi
th
the
c
hos
e
n
model.
T
his
dis
c
ove
r
y
pr
ovides
c
onc
lus
ive
e
videnc
e
that
thi
s
phe
n
omenon
is
l
inked
to
c
ha
nge
s
in
the
c
las
s
if
ica
ti
on
pa
r
t
o
f
f
ine
-
tuni
ng,
whic
h
c
a
n
in
f
luenc
e
the
model's
pe
r
f
or
manc
e
a
nd
c
ompl
e
xit
y
.
4.
CONC
L
USI
ON
T
hi
s
s
tu
dy
i
nt
e
g
r
a
te
s
f
in
e
-
t
un
in
g
B
E
R
T
w
it
h
R
N
N
s
f
or
pe
r
s
on
a
l
it
y
d
e
t
e
c
ti
on
u
s
in
g
X
d
a
t
a
.
T
he
R
N
N
s
u
s
e
d
c
on
s
i
s
t
of
v
a
ni
ll
a
R
NN
,
L
S
T
M
,
a
n
d
G
R
U.
I
n
a
dd
it
i
on,
t
hi
s
s
tu
dy
a
l
s
o
u
s
e
s
th
e
B
E
R
T
-
b
a
s
e
d
f
ul
ly
c
on
n
e
c
te
d
l
a
y
e
r
a
s
a
c
om
pa
r
i
s
on.
T
he
r
e
s
u
lt
s
i
nd
ic
a
t
e
t
ha
t
t
he
B
E
R
T
-
b
a
s
e
d
f
ull
y
c
o
nn
e
c
t
e
d
l
a
y
e
r
a
c
hi
e
ve
s
t
h
e
hi
gh
e
s
t
a
c
c
ur
a
c
y
f
or
c
l
a
s
s
I
/E
a
nd
F
/T
,
w
it
h
s
c
or
e
s
o
f
0.
5
62
a
n
d
0.
53
8.
T
h
i
s
i
s
a
tt
r
ib
ut
e
d
t
o
it
s
a
b
il
it
y
t
o
e
f
f
e
c
ti
v
e
l
y
c
a
pt
ur
e
f
r
e
q
u
e
n
tl
y
a
p
pe
a
r
in
g
wor
d
s
a
n
d
un
de
r
s
t
a
n
d
r
e
l
a
ti
on
s
hi
p
s
d
ir
e
c
tl
y
w
it
ho
ut
r
e
l
yi
ng
on
a
ti
m
e
s
e
q
u
e
n
c
e
.
O
n
th
e
o
th
e
r
ha
nd,
th
e
c
l
a
s
s
e
s
N
/S
a
n
d
J
/P
r
e
q
uir
e
t
o
u
n
de
r
s
ta
nd
mor
e
c
o
mpl
e
x
s
e
nt
e
n
c
e
s
tr
uc
tur
e
s
a
nd
m
a
n
a
g
e
t
e
m
por
a
l
s
e
qu
e
n
c
e
s
.
T
h
u
s
,
t
he
be
s
t
-
p
e
r
f
or
m
in
g
m
e
t
ho
d
f
or
t
hi
s
c
la
s
s
i
s
B
E
R
T
+
L
S
T
M
,
wit
h
a
c
c
ur
a
c
y
s
c
or
e
s
o
f
0.
5
43
a
nd
0.
53
2
.
T
hi
s
s
u
gge
s
t
s
th
a
t
i
nt
e
g
r
a
ti
on
wi
th
R
N
N
s
c
a
n
h
a
ve
a
p
o
s
it
iv
e
l
y
im
pa
c
t
o
n
c
e
r
ta
in
c
la
s
s
e
s
.
F
or
f
u
tur
e
wor
k,
s
e
v
e
r
a
l
i
mpr
ov
e
me
nt
s
c
a
n
b
e
e
xp
lor
e
d.
F
i
r
s
t,
op
ti
mi
z
i
ng
hy
p
e
r
p
a
r
a
m
e
t
e
r
s
e
tt
i
ng
s
i
s
e
s
s
e
n
ti
a
l,
a
s
R
N
N
s
a
r
e
hi
ghl
y
s
e
n
s
it
i
ve
t
o
th
e
hy
p
e
r
p
a
r
a
m
e
t
e
r
s
u
s
e
d.
T
e
c
hni
qu
e
s
s
u
c
h
a
s
B
a
ye
s
ia
n
op
ti
mi
z
a
ti
on
c
a
n
he
lp
id
e
nti
f
y
op
ti
ma
l
c
o
nf
i
gur
a
t
io
n
s
.
Ad
dit
io
n
a
ll
y,
i
t
ma
y
b
e
b
e
n
e
f
i
c
ia
l
to
a
lt
e
r
t
h
e
s
tr
u
c
tu
r
e
of
th
e
tw
e
e
t
da
t
a
b
y
s
e
gm
e
nti
ng
e
a
c
h
t
w
e
e
t
f
or
u
s
e
wi
th
B
E
R
T
.
M
a
int
a
ini
ng
t
he
a
u
th
e
nti
c
i
ty
of
e
a
c
h
tw
e
e
t
'
s
c
o
n
t
e
n
t
i
s
vi
ta
l
t
o
a
v
oid
m
ixi
ng
i
nf
or
m
a
t
io
n.
I
nv
e
s
ti
g
a
ti
ng
m
e
t
ho
d
s
l
ik
e
hi
e
r
a
r
c
h
ic
a
l
B
E
R
T
a
r
c
hi
te
c
t
ur
e
s
c
o
ul
d
a
l
s
o
b
e
v
a
lu
a
bl
e
in
c
a
pt
ur
i
ng
th
e
s
tr
u
c
tu
r
a
l
n
ua
n
c
e
s
of
t
h
e
t
e
xt
.
AC
KNOWL
E
DGE
M
E
NT
S
T
he
a
uthor
s
would
g
r
a
tef
ul
to
the
a
nonymous
r
e
vi
e
we
r
s
f
or
their
ins
ight
f
ul
c
omm
e
nts
,
c
ons
tr
uc
ti
ve
f
e
e
dba
c
k,
a
nd
he
lpf
ul
s
ugge
s
ti
ons
,
whic
h
gr
e
a
tl
y
e
nha
nc
e
d
the
qua
li
ty
of
thi
s
pa
pe
r
.
F
UN
DI
NG
I
NF
ORM
AT
I
ON
T
his
s
tudy
wa
s
f
unde
d
by
a
gr
a
nt
f
r
om
Di
r
e
c
t
or
a
te
of
R
e
s
e
a
r
c
h,
T
e
c
hnology
a
nd
C
omm
unit
y,
M
ini
s
tr
y
of
E
duc
a
ti
on
,
C
u
lt
ur
e
,
R
e
s
e
a
r
c
h,
a
nd
T
e
c
hnology,
I
ndone
s
ia
(
P
T
M
G
r
a
nt
449A
-
70/UN7.
D2/P
P
/VI
/2023
,
20
J
une
2023)
.
AU
T
HO
R
CONT
RI
B
U
T
I
ONS
S
T
AT
E
M
E
N
T
T
his
jour
na
l
us
e
s
the
C
ontr
ibut
o
r
R
oles
T
a
xo
nomy
(
C
R
e
diT
)
to
r
e
c
ognize
indi
vidual
a
uthor
c
ontr
ibut
ions
,
r
e
duc
e
a
utho
r
s
hip
dis
pu
tes
,
a
nd
f
a
c
il
it
a
te
c
oll
a
bor
a
ti
on.
Nam
e
of
Au
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
S
e
l
v
i
F
i
t
r
ia
K
h
oe
r
u
n
n
is
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
B
a
yu
S
ur
a
r
s
o
✓
✓
✓
✓
✓
✓
✓
R
e
tno
Kus
umaningr
um
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
li
z
a
ti
on
M
:
M
e
th
odol
ogy
So
:
So
f
twa
r
e
Va
:
Va
li
da
ti
on
Fo
:
Fo
r
ma
l
a
na
ly
s
is
I
:
I
nve
s
ti
ga
ti
on
R
:
R
e
s
our
c
e
s
D
:
D
a
ta
C
ur
a
ti
on
O
:
W
r
it
in
g
-
O
r
ig
in
a
l
D
r
a
f
t
E
:
W
r
it
in
g
-
R
e
vi
e
w
&
E
di
ti
ng
Vi
:
Vi
s
ua
li
z
a
ti
on
Su
:
Su
pe
r
vi
s
io
n
P
:
P
r
oj
e
c
t
a
dmi
ni
s
tr
a
ti
on
Fu
:
Fu
ndi
ng a
c
qui
s
it
io
n
CONF
L
I
CT
OF
I
NT
E
RE
S
T
S
T
AT
E
M
E
N
T
T
he
a
uthor
s
de
c
lar
e
that
ther
e
is
no
c
onf
li
c
t
of
int
e
r
e
s
t
a
s
s
oc
iate
d
with
thi
s
pa
pe
r
.
DA
T
A
AV
AI
L
A
B
I
L
I
T
Y
R
e
leva
nt
da
ta
a
r
e
a
va
il
a
ble
upon
r
e
que
s
t,
with
a
pp
r
opr
iate
pe
r
mi
s
s
ions
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
339
5
-
3403
3402
RE
F
E
RE
NC
E
S
[
1]
J
.
C
a
r
de
n,
R
.
J
.
J
one
s
,
a
nd
J
.
P
a
s
s
mor
e
,
“
D
e
f
in
in
g
s
e
lf
-
a
w
a
r
e
ne
s
s
in
th
e
c
ont
e
xt
of
a
dul
t
de
ve
lo
pme
nt
:
a
s
ys
te
ma
ti
c
li
te
r
a
tu
r
e
r
e
vi
e
w
,”
J
our
nal
of
M
anage
m
e
nt
E
duc
at
io
n
, vol
. 46, no. 1, pp.
140
–
177, F
e
b. 2022, doi:
10.1177/105256292
1990065.
[
2]
A
.
B
.
B
a
kke
r
a
nd
M
.
V.
W
oe
r
kom,
“
S
tr
e
ngt
hs
u
s
e
in
or
ga
n
iz
a
ti
ons
:
A
po
s
it
iv
e
a
ppr
oa
c
h
of
o
c
c
upa
ti
ona
l
h
e
a
lt
h,”
C
ana
di
an
P
s
y
c
hol
ogy
, vol
. 59, no. 1, pp. 38
–
46, 2018, doi:
10.1037/ca
p0
000120.
[
3]
A
.
E
.
M
.
va
n
V
ia
ne
n,
“
P
e
r
s
on
–
e
nvi
r
onme
nt
f
it
:
a
r
e
vi
e
w
of
i
t
s
ba
s
ic
te
ne
ts
,”
A
nnual
R
e
v
ie
w
of
O
r
gani
z
at
io
nal
P
s
y
c
hol
ogy
and
O
r
gani
z
at
io
nal
B
e
hav
io
r
, vol
. 5, no. 1, pp. 75
–
101, 2018, doi:
10.1146/a
nnur
e
v
-
or
gps
yc
h
-
032117
-
104702.
[
4]
R
. M
. H
e
r
r
, A
. E
.
M
.
V.
V
ia
ne
n, C
. B
os
le
, a
nd J
. E
. F
is
c
he
r
, “
P
e
r
s
ona
li
ty
t
ype
ma
tt
e
r
s
:
P
e
r
c
e
pt
io
ns
of
j
ob
de
ma
nds
, j
ob r
e
s
our
c
e
s
,
a
nd
th
e
ir
a
s
s
oc
ia
ti
ons
w
it
h
w
or
k
e
nga
ge
me
nt
a
nd
me
nt
a
l
he
a
lt
h,”
C
ur
r
e
nt
P
s
y
c
hol
ogy
,
vol
.
42,
no.
4,
pp.
2576
–
2590,
2023,
doi
:
10.1007/s
12144
-
021
-
01517
-
w.
[
5]
I
. B
. M
ye
r
s
,
G
if
ts
di
ff
e
r
in
g unde
r
s
ta
ndi
ng pe
r
s
onal
it
y
t
y
p
e
. M
o
unt
a
in
V
ie
w
, C
a
li
f
or
ni
a
:
D
a
vi
e
s
B
la
c
k P
ubl
is
hi
ng, 1995.
[
6]
N
.
H
.
J
e
r
e
my
a
nd
D
.
S
uha
r
to
no,
“
A
ut
om
a
ti
c
pe
r
s
on
a
li
ty
pr
e
di
c
ti
on
f
r
om
I
ndone
s
ia
n
us
e
r
on
twi
tt
e
r
u
s
in
g
w
or
d
e
mb
e
ddi
ng
a
nd
ne
ur
a
l
ne
twor
ks
,”
P
r
oc
e
di
a C
om
put
e
r
Sc
ie
n
c
e
, vol
. 179, pp. 41
6
–
422, 2021, doi:
10.1016/j
.pr
oc
s
.2021.01.024.
[
7]
M
.
F
r
kovi
ć
,
N
.
Č
e
r
ke
z
,
B
.
V
r
dol
ja
k,
a
nd
S
.
S
ka
ns
i,
“
E
va
lu
a
ti
on
of
s
tr
uc
tu
r
a
l
hype
r
pa
r
a
me
te
r
s
f
or
te
xt
c
la
s
s
if
ic
a
ti
on
w
it
h
L
S
T
M
ne
twor
ks
,”
in
2020
43r
d
I
nt
e
r
nat
io
nal
C
onv
e
nt
io
n
on
I
nf
or
m
a
ti
on,
C
om
m
uni
c
at
io
n
and
E
le
c
tr
oni
c
T
e
c
hnol
ogy
(
M
I
P
R
O
)
,
20
20,
pp. 145
–
150
, doi
:
10.23919/M
I
P
R
O
48935.2020.9245216.
[
8]
G
. R
ya
n, P
. K
a
ta
r
in
a
, a
nd D
. S
uha
r
to
no, “
M
B
T
I
pe
r
s
ona
li
ty
pr
e
di
c
ti
on us
in
g ma
c
hi
ne
l
e
a
r
ni
ng a
nd s
mot
e
f
or
ba
la
nc
in
g da
ta
ba
s
e
d
on s
ta
te
me
nt
s
e
nt
e
nc
e
s
,
”
I
nf
or
m
at
io
n
, vol
. 14, no. 4, 2023, doi
:
10.3390/i
nf
o14040217.
[
9]
N
.
A
ga
r
w
a
l
e
t
al
.
,
“
P
e
r
s
on
a
li
ty
pr
e
di
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
us
in
g
T
w
it
te
r
da
ta
,
”
Soc
ia
l
N
e
tw
or
k
in
g
and
C
om
put
at
i
onal
I
nt
e
ll
ig
e
nc
e
,
pp. 707
–
716, 2020, doi:
10.1007/978
-
981
-
15
-
2071
-
6_59.
[
10]
K
.
A
.
N
is
ha
,
U
.
K
ul
s
um,
S
.
R
a
hma
n,
M
d.
F
.
H
os
s
a
in
,
P
.
C
ha
kr
a
bor
ty
,
a
nd
T
.
C
houdhur
y,
“
A
c
ompa
r
a
ti
ve
a
na
ly
s
is
of
ma
c
hi
ne
le
a
r
ni
ng
a
ppr
oa
c
he
s
in
p
e
r
s
ona
li
ty
pr
e
di
c
ti
on
us
in
g
M
B
T
I
,”
i
n
C
om
put
at
io
nal
I
nt
e
ll
ig
e
nc
e
in
P
at
te
r
n
R
e
c
ogni
ti
on
,
S
in
ga
por
e
:
S
pr
in
ge
r
,
2022, pp. 13
–
23
, doi
:
10.1007/978
-
981
-
16
-
2543
-
5_2.
[
11]
P
.
F
.
M
u
h
a
mm
a
d,
R
.
K
u
s
u
ma
ni
ng
r
um
,
a
n
d
A
.
W
i
bo
w
o,
“
S
e
n
ti
m
e
nt
a
n
a
ly
s
i
s
u
s
i
ng
W
or
d
2v
e
c
a
n
d
l
on
g
s
hor
t
-
t
e
r
m
m
e
m
or
y
(
L
S
T
M
)
f
or
I
n
do
ne
s
ia
n
ho
te
l
r
e
v
i
e
w
s
,
”
P
r
oc
e
di
a
C
o
m
put
e
r
S
c
i
e
nc
e
,
vol
.
17
9
,
pp
.
72
8
–
73
5,
2
02
1,
d
oi
:
1
0.
10
16
/j
.pr
oc
s
.2
02
1.
01
.0
61
.
[
12]
H
.
N
a
ik
,
S
.
D
e
dhi
a
,
A
.
D
ubbe
w
a
r
,
M
.
J
os
hi
,
a
nd
V
.
P
a
ti
l,
“
M
ye
r
s
B
r
ig
gs
ty
pe
in
di
c
a
to
r
(
M
B
T
I
)
-
pe
r
s
ona
li
ty
pr
e
di
c
ti
on
us
in
g
de
e
p
le
a
r
ni
ng
,”
in
2022
2nd
A
s
ia
n
C
onf
e
r
e
nc
e
on
I
n
nov
at
io
n
in
T
e
c
hnol
ogy
(
A
SI
A
N
C
O
N
)
,
2022,
pp.
1
–
6
,
doi
:
10.1109/AS
I
A
N
C
O
N
55314.2022.9909077.
[
13]
M
.
M
a
ul
id
a
h
a
nd
H
.
F
.
P
a
r
de
de
,
“
P
r
e
di
c
ti
on
of
M
ye
r
s
-
B
r
ig
gs
ty
pe
in
di
c
a
to
r
pe
r
s
ona
li
ty
us
in
g
lo
ng
s
hor
t
-
te
r
m
me
mor
y
,”
J
ur
nal
E
le
k
tr
oni
k
a dan T
e
le
k
om
uni
k
as
i
, vol
. 21, no. 2, pp. 104, 2021,
doi
:
10.14203/j
e
t.
v21.104
-
111.
[
14]
J
.
D
e
vl
in
,
M
.
-
W
.
C
h
a
ng,
K
.
L
e
e
,
a
nd
K
.
T
out
a
nova
,
“
B
E
R
T
:
pr
e
-
tr
a
in
in
g
of
de
e
p
bi
di
r
e
c
ti
ona
l
tr
a
ns
f
or
me
r
s
f
or
la
ng
ua
ge
unde
r
s
ta
ndi
ng
,”
in
P
r
oc
e
e
di
ngs
of
th
e
2
019
C
onf
e
r
e
nc
e
of
th
e
N
or
th
A
m
e
r
ic
an
C
hapt
e
r
of
th
e
A
s
s
oc
ia
ti
on
fo
r
C
om
put
at
i
onal
L
in
gui
s
ti
c
s
:
H
um
an L
anguage
T
e
c
hnol
ogi
e
s
, 2019, pp. 4171
–
4
176.
[
15]
H
.
Z
ha
ng,
“
M
B
T
I
pe
r
s
ona
li
ty
pr
e
di
c
ti
on
ba
s
e
d
on
B
E
R
T
c
la
s
s
if
ic
a
ti
on
”
,
in
4t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
C
om
put
e
r
Sc
ie
nc
e
and I
nt
e
ll
ig
e
nt
C
om
m
uni
c
at
io
n (
C
SI
C
2022)
, vol
.34, 2023, doi:
10.54097/hs
e
t.
v34i
.5497
[
16]
Z
.
R
e
n,
Q
.
S
he
n,
X
.
D
ia
o,
a
nd
H
.
X
u,
“
A
s
e
nt
im
e
nt
-
a
w
a
r
e
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
h
f
or
pe
r
s
ona
li
ty
d
e
te
c
ti
on
f
r
om
te
xt
,”
I
nf
or
m
at
io
n P
r
oc
e
s
s
in
g & M
anage
m
e
nt
, vol
. 58, no. 3, 2021, d
oi
:
10.1016/j
.i
pm.2021.102532.
[
17]
V
.
G
.
D.
S
a
nt
os
a
nd
I
.
P
a
r
a
boni
,
“
M
ye
r
s
-
B
r
ig
gs
p
e
r
s
ona
li
ty
c
la
s
s
if
ic
a
ti
on
f
r
om
s
oc
ia
l
m
e
di
a
te
xt
us
in
g
pr
e
-
tr
a
in
e
d
la
ng
ua
ge
mode
ls
,”
J
U
C
S
-
J
ou
r
nal
of
U
ni
v
e
r
s
al
C
om
put
e
r
Sc
ie
nc
e
, vol
.
28,
no. 4, pp. 378
-
395, 2022, doi:
10.3897/j
uc
s
.70941.
[
18]
S
.
S
.
K
e
h
a
nd
I
.
-
T
.
C
he
ng,
“
M
ye
r
s
-
B
r
ig
gs
pe
r
s
ona
li
ty
c
la
s
s
if
ic
a
ti
on
a
nd
pe
r
s
ona
li
ty
-
s
pe
c
if
ic
la
ngua
g
e
ge
ne
r
a
ti
on
u
s
in
g
pr
e
-
tr
a
in
e
d l
a
ngua
ge
mod
e
ls
,”
a
r
X
iv
-
C
om
put
e
r
Sc
ie
n
c
e
,
pp. 1
-
5,
20
19, doi:
10.4855
0/
a
r
X
iv
.1907.06333.
[
19]
A
.
N
.
A
z
ha
r
,
“
F
in
e
-
tu
ni
ng
p
r
e
tr
a
in
e
d
mul
ti
li
ngua
l
B
E
R
T
mode
l
f
or
I
ndone
s
ia
n
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,
”
2020
7t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
A
dv
anc
e
d
I
nf
or
m
at
ic
s
:
C
on
c
e
pt
s
,
T
he
or
y
and
A
ppl
ic
at
io
ns
,
I
C
A
I
C
T
A
2020
,
2020,
doi
:
10
.1109/I
C
A
I
C
T
A
49861.2020.9428882.
[
20]
L
.
Z
hu
a
nd
D
.
L
uo,
“
A
nove
l
e
f
f
ic
ie
nt
a
nd
e
f
f
e
c
ti
ve
pr
e
pr
oc
e
s
s
in
g
a
lg
or
it
hm
f
or
te
xt
c
la
s
s
if
ic
a
ti
on
,”
J
our
nal
of
C
om
put
e
r
and
C
om
m
uni
c
at
io
ns
, vol
. 11, no. 03, pp. 1
–
14, 2023, doi:
10.4236/j
c
c
.2023.113001.
[
21]
T
.
H
.
S
a
put
r
o
a
nd
A
.
H
e
r
ma
w
a
n,
“
T
he
a
c
c
ur
a
c
y
im
pr
ove
me
nt
of
te
xt
mi
ni
ng
c
la
s
s
if
ic
a
ti
on
on
hos
pi
ta
l
r
e
vi
e
w
th
r
ough
th
e
a
lt
e
r
a
ti
on
in
th
e
pr
e
pr
oc
e
s
s
in
g
s
ta
ge
,
”
I
nt
e
r
nat
io
nal
J
our
n
al
of
C
om
put
e
r
and
I
nf
or
m
at
io
n
T
e
c
hnol
ogy
,
vol
.
10,
no.
4,
pp. 2279
–
0764, 2021, doi:
1
0.24203/i
jc
it
.v10i4.138.
[
22]
B
.
A
.
H
.
M
ur
s
h
e
d,
S
.
M
a
ll
a
pp
a
,
O
.
A
.
M
.
G
ha
le
b,
a
nd
H
.
D
.
E
.
A
l
-
a
r
ik
i,
“
E
f
f
ic
ie
nt
T
w
it
te
r
da
ta
c
le
a
ns
in
g
mode
l
f
or
da
ta
a
na
l
ys
i
s
of
th
e
pa
nde
m
ic
T
w
e
e
ts
,
”
in
St
udi
e
s
in
Sy
s
te
m
s
,
D
e
c
i
s
io
n
and
C
ont
r
ol
,
vol
.
348,
2021,
pp.
93
–
114
,
doi
:
10.1007/978
-
3
-
0
30
-
67716
-
9_7.
[
23]
A
.
L
.
R
io
,
M
.
M
a
r
ti
n,
A
.
P
e
r
e
r
a
-
L
lu
na
,
a
nd
R
.
S
a
id
i,
“
E
f
f
e
c
t
o
f
s
e
que
nc
e
pa
ddi
ng
on
th
e
pe
r
f
or
ma
nc
e
of
de
e
p
le
a
r
ni
ng
mode
l
s
in
a
r
c
ha
e
a
l
pr
ot
e
in
f
unc
ti
ona
l
pr
e
di
c
ti
on,”
Sc
ie
nt
if
ic
R
e
por
ts
, vol
.
10, no. 1, pp. 1
–
14, 20
20, doi:
10.1038/s
41598
-
020
-
71450
-
8.
[
24]
D
.
G
.
M
a
ndha
s
iy
a
,
H
.
M
ur
f
i,
a
nd
A
.
B
us
ta
ma
m,
“
T
he
hybr
id
of
B
E
R
T
a
nd
de
e
p
le
a
r
ni
ng
mode
ls
f
or
I
ndone
s
ia
n
s
e
nt
im
e
nt
a
na
ly
s
is
,”
I
ndone
s
ia
n
J
ou
r
nal
of
E
le
c
tr
ic
al
E
ngi
ne
e
r
in
g
and
C
om
put
e
r
Sc
i
e
nc
e
,
vol
.
33,
no.
1,
pp.
5
91
–
602,
J
a
n.
2024,
doi
:
10.11591/i
je
e
c
s
.v33.i1.pp591
-
602.
[
25]
K
.
S
.
R
a
o,
D
.
V
a
ll
ur
u,
S
.
P
a
tn
a
la
,
R
.
B
.
D
e
va
r
e
ddi
,
T
.
S
.
R
.
K
r
is
hna
,
a
nd
A
.
S
r
a
va
ni
,
“
P
hi
s
hi
ng
w
e
b
s
it
e
de
te
c
ti
on
u
s
in
g
n
ove
l
in
te
gr
a
ti
on
of
B
E
R
T
a
nd
X
L
N
e
t
w
it
h
de
e
p
le
a
r
ni
ng
s
e
que
n
ti
a
l
mode
ls
,
”
I
ndone
s
ia
n
J
our
nal
of
E
le
c
tr
ic
al
E
ngi
ne
e
r
in
g
and
C
om
put
e
r
Sc
ie
n
c
e
, vol
. 36, no. 2, p
p
. 1273
-
1283
, N
ov. 2024, d
oi
:
10.11591/i
je
e
c
s
.v36.i2.pp1273
-
1283.
[
26]
K
.
S
.
N
ugr
oho,
A
.
Y.
S
ukma
de
w
a
,
D
.
W
.
H
.
W
us
w
il
a
ha
ke
n
,
F
.
A
.
B
a
c
ht
ia
r
,
a
nd
N
.
Y
udi
s
ti
r
a
,
“
B
E
R
T
f
in
e
-
tu
ni
ng
f
or
s
e
nt
i
me
nt
a
na
ly
s
is
on
in
done
s
ia
n
mobi
le
a
pps
r
e
vi
e
w
s
,
”
in
SI
E
T
’
21:
P
r
oc
e
e
di
ngs
of
th
e
6t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
Sus
ta
in
abl
e
I
nf
or
m
at
io
n E
ngi
ne
e
r
in
g and T
e
c
hnol
ogy
, 2021, pp. 258
–
264
,
doi
:
10.1145/3479645.34
79679.
[
27]
N
.
J
.
J
oha
nne
s
e
n,
M
.
L
.
K
ol
he
,
a
nd
M
.
G
oodwin,
“
C
ompa
r
in
g
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
us
in
g
pr
in
c
ip
a
l
c
ompone
nt
a
na
ly
s
is
f
or
e
le
c
tr
ic
a
l
lo
a
d
pr
e
di
c
ti
ons
,”
in
2021
6t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
Sm
ar
t
and
Sus
ta
in
abl
e
T
e
c
hnol
ogi
e
s
(
Spl
iT
e
c
h)
,
2021,
pp.
1
–
6
, d
oi
:
10.23919/S
pl
iT
e
c
h52315.2021.9566357.
[
28]
M
.
A
.
R
iz
a
a
nd
N
.
C
ha
r
ib
a
ld
i,
“
E
mot
io
n
de
te
c
ti
on
in
T
w
it
te
r
s
oc
ia
l
me
di
a
u
s
in
g
lo
ng
s
hor
t
-
te
r
m
me
mor
y
(
L
S
T
M
)
a
nd
f
a
s
t
te
xt
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
& R
obot
ic
s
(
I
J
A
I
R
)
, vol
. 3, no. 1, p
p. 15
–
26, 2021, doi:
10.25139/i
ja
i
r
.v3i
1.3827
.
[
29]
J
.
S
hi
,
S
.
W
a
ng,
P
.
Q
u,
a
nd
J
.
S
ha
o,
“
T
im
e
s
e
r
ie
s
pr
e
di
c
ti
on
mode
l
us
in
g
L
S
T
M
-
t
r
a
ns
f
or
me
r
ne
ur
a
l
ne
twor
k
f
or
mi
ne
w
a
te
r
in
f
lo
w
,”
Sc
ie
nt
if
ic
R
e
por
ts
, vol
. 14, 2024, doi:
10.1038/s
41598
-
024
-
69418
-
z.
[
30]
E.
E
z
hi
la
r
a
s
i
I
a
nd
J
.
C
.
C
le
me
nt
,
“
G
R
U
-
S
V
M
ba
s
e
d
th
r
e
a
t
d
e
te
c
ti
on
in
c
ogni
ti
ve
r
a
di
o
ne
twor
k
,”
Se
ns
or
s
,
vol
.
23,
no.
3,
F
e
b.
2023, doi:
10.3390/s
23031326.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
F
ine
-
tuni
ng
b
idi
r
e
c
ti
onal
e
nc
ode
r
r
e
pr
e
s
e
ntat
ions
fr
om
tr
ans
for
me
r
s
for
…
(
Se
lvi
F
it
r
ia
K
hoe
r
unnis
a)
3403
[
31]
Y
.
X
ia
o,
C
.
Z
ou,
H
.
C
hi
,
a
nd
R
.
F
a
ng,
“
B
oo
s
te
d
G
R
U
mode
l
f
or
s
hor
t
-
te
r
m
f
o
r
e
c
a
s
ti
ng
of
w
in
d
pow
e
r
w
it
h
f
e
a
tu
r
e
-
w
e
ig
ht
e
d
pr
in
c
ip
a
l
c
ompone
nt
a
na
ly
s
is
,”
E
ne
r
gy
, vol
. 267, 2023, doi:
10.
1016/j
.e
ne
r
gy.2022.126503.
[
32]
F
.
K
ot
o,
A
.
R
a
hi
mi
,
J
.
H
.
L
a
u,
a
nd
T
.
B
a
ld
w
in
,
“
I
ndoL
E
M
a
nd
I
ndoB
E
R
T
:
A
be
nc
hma
r
k
da
ta
s
e
t
a
nd
pr
e
-
tr
a
in
e
d
la
ngua
ge
mo
de
l
f
or
I
ndone
s
ia
n
N
L
P
,”
in
P
r
oc
e
e
di
ngs
of
th
e
28t
h
I
n
te
r
nat
io
nal
C
onf
e
r
e
nc
e
on
C
om
put
at
io
nal
L
in
gui
s
ti
c
s
,
B
a
r
c
e
lo
na
,
S
pa
in
,
N
ov.
2020, pp. 757
–
770.
[
33]
A
.
D
a
tt
a
,
S
.
C
ha
kr
a
bor
ty
,
a
nd
A
.
M
ukhe
r
je
e
,
“
P
e
r
s
on
a
li
ty
de
te
c
ti
on
a
nd
a
na
ly
s
i
s
us
in
g
T
w
it
te
r
da
ta
,”
a
r
X
iv
-
C
om
put
e
r
Sc
i
e
nc
e
,
pp. 1
-
9,
2023, doi:
10.48550/ar
X
iv
.2309.05497.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
Sel
v
i
F
i
tr
i
a
Kho
erunn
i
s
a
recei
v
e
d
h
er
B.
S.
d
eg
re
e
(cu
ml
au
d
e)
i
n
i
n
f
o
rmat
i
cs
fro
m
U
n
i
v
ers
i
t
a
s
D
i
p
o
n
eg
o
ro
,
Semaran
g
,
In
d
o
n
es
i
a,
i
n
2
0
2
2
.
N
o
w
,
s
h
e
i
s
p
o
s
t
g
ra
d
u
a
t
e
s
t
u
d
en
t
at
Mas
t
er
o
f
In
fo
rma
t
i
o
n
Sy
s
t
em,
Sch
o
o
l
o
f
Po
s
t
g
rad
u
at
es
St
u
d
y
,
U
n
i
v
er
s
i
t
as
D
i
p
o
n
eg
o
ro
.
Sh
e
s
erv
e
d
as
a
L
ab
o
ra
t
o
r
y
A
s
s
i
s
t
a
n
t
w
i
t
h
t
h
e
D
e
p
art
me
n
t
o
f
In
fo
rmat
i
cs
,
U
n
i
v
ers
i
t
a
s
D
i
p
o
n
e
g
o
r
o
,
fro
m
2
0
2
2
t
o
n
o
w
.
Sh
e
i
s
cu
rre
n
t
l
y
a
Re
s
earch
A
s
s
i
s
t
an
t
w
i
t
h
t
h
e
L
ab
o
rat
o
ry
o
f
In
t
el
l
i
g
en
t
Sy
s
t
em
s
,
D
ep
ar
t
men
t
o
f
In
fo
rmat
i
cs
,
U
n
i
v
ers
i
t
a
s
D
i
p
o
n
eg
o
ro
.
H
er
res
earch
i
n
t
ere
s
t
s
i
n
cl
u
d
e
m
ach
i
n
e
l
ear
n
i
n
g
a
n
d
n
at
u
ral
l
an
g
u
a
g
e
p
r
o
ces
s
i
n
g
.
Sh
e
can
b
e
c
o
n
t
act
e
d
at
ema
i
l
:
s
el
v
i
f
k
h
@
s
t
u
d
en
t
s
.
u
n
d
i
p
.
ac.
i
d
.
Ba
y
u
Sura
rs
o
recei
v
ed
h
i
s
B.
S.
d
eg
ree
i
n
mat
h
ema
t
i
c
s
fro
m
U
n
i
v
er
s
i
t
as
G
ad
j
a
h
Mad
a,
Y
o
g
y
a
k
art
a,
I
n
d
o
n
e
s
i
a,
i
n
1
9
8
7
.
H
i
s
M.
S.
an
d
Ph
.
D
.
d
e
g
rees
fr
o
m
H
i
ro
s
h
i
ma
U
n
i
v
er
s
i
t
y
,
J
a
p
an
i
n
1
9
9
5
an
d
1
9
9
8
,
res
p
ec
t
i
v
el
y
.
H
e
i
s
cu
rren
t
l
y
a
l
ect
u
rer
at
t
h
e
D
e
p
art
me
n
t
o
f
Mat
h
emat
i
cs
,
Fac
u
l
t
y
o
f
Sci
e
n
ce
an
d
Ma
t
h
ema
t
i
c
s
,
U
n
i
v
er
s
i
t
as
D
i
p
o
n
eg
o
ro
.
H
er
res
earc
h
i
n
t
eres
t
s
i
n
cl
u
d
e
a
l
g
e
b
ra,
co
mb
i
n
a
t
o
r
i
cs
,
an
d
mat
h
emat
i
cal
l
o
g
i
c.
H
e
i
s
al
s
o
i
n
v
o
l
v
e
d
i
n
e
x
p
er
t
s
y
s
t
em
ap
p
l
i
ca
t
i
o
n
u
s
ed
t
o
i
d
e
n
t
i
fy
i
s
s
u
e
s
an
d
al
t
ern
at
i
v
e
s
o
l
u
t
i
o
n
s
fo
r
s
ec
o
n
d
ary
s
c
h
o
o
l
s
t
u
d
e
n
t
s
.
H
e
can
b
e
co
n
t
ac
t
ed
a
t
emai
l
:
b
a
y
u
s
@
l
ect
u
rer.
u
n
d
i
p
.
ac.
i
d
.
R
etno
Kus
u
m
a
n
i
ng
ru
m
recei
v
ed
h
er
B.
S.
d
eg
ree
i
n
mat
h
emat
i
cs
fr
o
m
U
n
i
v
er
s
i
t
as
D
i
p
o
n
e
g
o
r
o
,
Semaran
g
,
In
d
o
n
es
i
a,
i
n
2
0
0
3
,
an
d
h
er
M.
S.
an
d
Ph
.
D
.
d
eg
ree
s
fro
m
U
n
i
v
er
s
i
t
as
In
d
o
n
es
i
a,
D
ep
o
k
,
In
d
o
n
es
i
a
i
n
2
0
1
0
an
d
2
0
1
4
,
res
p
ec
t
i
v
el
y
.
Sh
e
i
s
cu
rren
t
l
y
a
l
ect
u
rer
at
t
h
e
D
e
p
art
me
n
t
o
f
I
n
fo
rma
t
i
c
s
,
Facu
l
t
y
o
f
Sci
en
ce
an
d
Mat
h
emat
i
cs
,
U
n
i
v
ers
i
t
a
s
D
i
p
o
n
eg
o
ro
.
Sh
e
al
s
o
cu
rre
n
t
l
y
s
er
v
es
a
s
t
h
e
h
ea
d
o
f
t
h
e
In
t
el
l
i
g
en
t
Sy
s
t
ems
L
ab
o
ra
t
o
ry
i
n
D
ep
ar
t
men
t
o
f
I
n
fo
rma
t
i
c
s
.
H
er
re
s
earch
i
n
t
ere
s
t
s
i
n
cl
u
d
e
n
at
u
ral
l
an
g
u
a
g
e
p
r
o
ces
s
i
n
g
,
mach
i
n
e
l
earn
i
n
g
,
co
mp
u
t
er
v
i
s
i
o
n
,
p
at
t
ern
reco
g
n
i
t
i
o
n
,
an
d
t
o
p
i
c
mo
d
e
l
i
n
g
.
Sh
e
i
s
a
memb
er
o
f
t
h
e
I
E
E
E
Co
mp
u
t
a
t
i
o
n
a
l
In
t
el
l
i
g
en
ce
So
c
i
et
y
,
IE
E
E
Co
mp
u
t
er
So
ci
e
t
y
,
an
d
A
CM.
H
er
aw
ard
s
an
d
h
o
n
o
r
s
i
n
cl
u
d
e
t
h
e
San
d
w
i
ch
-
L
i
k
e
s
ch
o
l
ar
s
h
i
p
aw
ar
d
fr
o
m
t
h
e
D
i
rec
t
o
ra
t
e
G
en
era
l
o
f
H
i
g
h
er
E
d
u
cat
i
o
n
o
f
I
n
d
o
n
e
s
i
a
fo
r
v
i
s
i
t
i
n
g
t
h
e
Sch
o
o
l
o
f
S
y
s
t
em
E
n
g
i
n
eeri
n
g
,
U
n
i
v
er
s
i
t
y
o
f
Read
i
n
g
,
Read
i
n
g
,
U
.
K
.
a
s
a
s
t
u
d
e
n
t
v
i
s
i
t
o
r
i
n
2
0
1
2
,
t
h
e
Bes
t
Pa
p
er
o
f
t
h
e
Sec
o
n
d
I
n
t
er
n
at
i
o
n
al
Co
n
fere
n
ce
o
n
In
fo
rma
t
i
c
s
an
d
Co
m
p
u
t
at
i
o
n
al
Sci
e
n
ce
s
i
n
2
0
1
8
,
fi
rs
t
p
l
ac
e
fo
r
O
u
t
s
t
a
n
d
i
n
g
L
ect
u
rer
-
U
n
i
v
ers
i
t
a
s
D
i
p
o
n
e
g
o
r
o
fo
r
t
h
e
Sci
e
n
ce
an
d
T
ech
n
o
l
o
g
y
Ca
t
eg
o
ry
i
n
2
0
1
9
,
an
d
s
eco
n
d
p
l
ace
fo
r
t
h
e
Be
s
t
Pa
p
er
A
w
ar
d
o
f
t
h
e
T
h
i
rd
I
n
t
er
n
at
i
o
n
al
Sy
m
p
o
s
i
u
m
o
n
A
d
v
a
n
ced
In
t
e
l
l
i
g
e
n
t
In
f
o
rmat
i
cs
i
n
2
0
2
0
.
Sh
e
can
b
e
c
o
n
t
act
e
d
at
emai
l
:
re
t
n
o
@
l
i
v
e.
u
n
d
i
p
.
ac.
i
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.