I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
2025
, pp.
641
~
649
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
1
.pp
641
-
649
641
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
C
h
i
n
e
se
p
ap
e
r
c
l
ass
i
f
i
c
at
i
on
b
ase
d
on
p
r
e
-
t
r
ai
n
e
d
l
an
gu
age
m
od
e
l
an
d
h
y
b
r
i
d
d
e
e
p
l
e
ar
n
i
n
g
m
e
t
h
od
X
in
L
u
o
, S
of
ia
n
it
a M
u
t
al
ib
,
S
yar
ip
ah
R
u
z
ai
n
i
S
ye
d
A
r
is
S
c
h
oo
l
o
f
C
om
pu
t
i
ng
S
c
i
e
n
c
e
s
,
C
ol
l
e
ge
of
C
o
m
pu
t
i
ng
,
I
n
f
o
r
m
a
t
i
c
s
a
nd
M
a
t
he
m
a
t
i
c
s
,
U
n
i
ve
r
s
i
t
i
T
e
k
no
l
og
i
M
A
R
A
,
S
ha
h
A
l
a
m
, M
a
l
a
ys
ia
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
y 6, 2024
R
e
vi
s
e
d
S
e
p 11, 2024
A
c
c
e
pt
e
d
S
e
p 30, 2024
With
the
explosive
growth
in
the
number
of
published
papers,
rese
archers
must filter pa
pers by c
ategory to
improve re
trieval
efficie
ncy. The f
eat
ures of
data
can
be
learned
through
complex
network
structures
of
deep
le
arning
models
without
the
need
for
man
ual
definition
and
extraction
in
ad
vance,
resulting
in better
processing
performance
for large d
atasets. In o
ur
stu
dy, the
pre
-
trained
language
model
bidirectional
encoder
representations
from
transforme
rs
(BERT)
and
other
deep
learning
models
were
applied
to
paper
classifi
cation.
A
large
-
scale
chinese
scientific
literature
dataset
was
used,
including
abstracts
,
keywords,
titles,
disciplines,
and
categor
ies
from
396
k
papers.
Currently,
there
is
little
in
-
depth
research
on
the
role
of
titles,
abstracts
,
and
keywords
in
classifi
cation
and
how
they
are
u
sed
in
combinat
ion.
To
address
this
issue,
we
evaluated
classifi
cation
res
ults
by
employi
ng
different
title,
abstract,
and
keywords
concatenation
meth
ods
to
generate
model
input
data,
and
compared
the
effects
o
f
a
single
sente
nce
or
sentence
pair
data
input
methods.
We
also
adopted
an
ensemble
le
arning
approach
to
integrat
e
the
results
of
models
that
processed
titles
,
keywor
ds,
and
abstracts
independen
tly
to
find
the
best
combinat
ion.
Finall
y,
we
studi
ed
the
com
bination
of
different
types
of
models,
such
as
the
combination
of
BERT
and
convolut
ional n
eural netwo
rks (CNN)
, and
measured the performa
nce by
accuracy, weig
hted average
precision
, weight
ed average recal
l, and w
e
ighted
average F1
score.
K
e
y
w
o
r
d
s
:
B
id
ir
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
C
hi
ne
s
e
s
c
i
e
nt
if
ic
li
te
r
a
tu
r
e
da
ta
s
e
t
D
e
e
p l
e
a
r
ni
ng mode
l
M
ode
l
c
om
bi
na
ti
on
P
a
pe
r
c
la
s
s
if
ic
a
ti
on
P
r
e
-
tr
a
in
in
g l
a
ngua
ge
m
ode
l
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
S
of
ia
ni
ta
M
ut
a
li
b
S
c
hool
of
C
om
put
in
g S
c
ie
nc
e
s
, C
ol
le
ge
of
C
om
put
in
g, I
nf
or
m
a
ti
c
s
a
nd M
a
th
e
m
a
ti
c
s
U
ni
ve
r
s
it
i
T
e
knol
ogi
M
A
R
A
S
ha
h A
la
m
, S
e
la
ngor
40450
,
M
a
la
ys
ia
E
m
a
il
:
s
of
ia
ni
ta
@
ui
tm
.e
du.my
1.
I
N
T
R
O
D
U
C
T
I
O
N
I
n r
e
c
e
nt
ye
a
r
s
, due
t
o
th
e
i
nc
r
e
a
s
in
g numbe
r
o
f
s
c
ie
nt
if
ic
pa
pe
r
s
, r
e
s
e
a
r
c
he
r
s
ne
e
d t
o r
e
tr
ie
ve
pa
pe
r
s
r
e
la
te
d
to
th
e
ir
r
e
s
e
a
r
c
h
f
ie
ld
s
m
or
e
e
f
f
ic
ie
nt
ly
.
T
he
c
a
te
gor
y
la
be
li
ng
of
s
c
ie
nt
if
ic
pa
pe
r
s
is
a
ta
s
k
th
a
t
m
us
t
be
c
om
pl
e
te
d i
n doc
um
e
nt
t
a
xonomy. I
f
it
i
s
c
om
pl
e
te
d by ma
npowe
r
, pr
of
e
s
s
io
na
l
knowle
dge
m
us
t
be
r
e
qui
r
e
d,
w
hi
c
h
is
c
os
tl
y
a
nd
in
e
f
f
ic
ie
nt
.
I
t
is
a
n
im
por
ta
nt
ta
s
k
in
na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
(
N
L
P
)
to
c
om
pl
e
te
th
e
a
ut
om
a
ti
c
c
la
s
s
if
ic
a
ti
on
of
pa
pe
r
s
th
r
ough
m
a
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm
s
a
nd
a
c
hi
e
ve
pr
a
c
ti
c
a
l
a
c
c
ur
a
c
y.
A
ppl
yi
ng
tr
a
di
ti
ona
l
m
a
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm
s
to
c
la
s
s
if
y
pa
pe
r
s
r
e
qui
r
e
s
c
om
pl
e
ti
ng
two
s
te
ps
.
F
ir
s
t,
obt
a
in
th
e
doc
um
e
nt
r
e
pr
e
s
e
nt
a
ti
on
v
e
c
to
r
th
r
ough
te
r
m
f
r
e
que
nc
y
-
in
ve
r
s
e
doc
um
e
nt
f
r
e
que
nc
y
(
TF
-
I
D
F
)
,
W
or
d2V
e
c
,
gl
oba
l
ve
c
to
r
s
f
or
w
or
d
r
e
pr
e
s
e
nt
a
ti
on (
G
lo
V
e
c
)
, F
a
s
tT
e
xt
,
a
nd othe
r
m
e
th
ods
, a
nd t
he
n us
e
t
he
m
a
s
in
put
da
t
a
f
or
c
la
s
s
if
ic
a
ti
on
a
lg
or
it
hm
s
s
uc
h
a
s
na
iv
e
B
a
ye
s
,
de
c
i
s
io
n
tr
e
e
,
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
,
a
nd
ne
ur
a
l
ne
twor
k
[
1]
–
[
4]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
20
25
:
641
-
649
642
T
r
a
di
ti
ona
l
m
a
c
hi
ne
le
a
r
ni
ng
c
la
s
s
if
ie
r
s
ha
v
e
hi
gh
a
c
c
ur
a
c
y
a
nd
e
f
f
ic
ie
nc
y
in
s
m
a
ll
da
ta
s
e
ts
,
but
s
tr
uggl
e
w
it
h
la
r
ge
-
s
c
a
le
da
ta
s
e
ts
w
it
h
c
om
pl
e
x
f
e
a
tu
r
e
s
.
D
e
e
p
l
e
a
r
ni
ng
of
f
e
r
s
a
dva
nt
a
ge
s
li
ke
a
voi
di
ng
m
a
nua
l
f
e
a
tu
r
e
de
f
in
it
io
n
a
nd
us
in
g
c
om
pl
e
x
ne
twor
k
s
tr
uc
tu
r
e
s
to
e
xt
r
a
c
t
a
nd
ge
ne
r
a
li
z
e
da
ta
f
e
a
tu
r
e
s
,
r
e
s
ul
ti
ng
in
hi
ghe
r
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y. C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
)
c
a
n e
xt
r
a
c
t
lo
c
a
l
f
e
a
tu
r
e
s
a
nd
tr
a
ns
f
e
r
th
e
m
to
gl
oba
l
f
e
a
tu
r
e
s
by
pool
in
g
la
ye
r
s
f
r
om
te
xt
s
e
qu
e
nc
e
s
[
5]
,
[
6]
.
R
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
k
(
R
N
N
)
a
nd
lo
ng
s
hor
t
-
te
r
m
m
e
m
or
y
(
L
S
T
M
)
m
ode
ls
a
r
e
s
ui
ta
bl
e
f
or
pr
oc
e
s
s
in
g
s
e
que
nt
ia
l
da
ta
b
e
c
a
u
s
e
th
e
y
c
a
n
r
e
m
e
m
be
r
th
e
de
pe
nde
nc
y
be
twe
e
n
to
ke
ns
[
7]
,
[
8]
.
B
id
ir
e
c
ti
o
na
l
e
nc
ode
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
(
B
E
R
T
)
is
a
bi
di
r
e
c
ti
ona
l
la
ngua
ge
m
ode
l
th
a
t
pe
r
f
or
m
s
m
a
s
k
la
ngua
ge
m
ode
l
(
M
L
M
)
ta
s
ks
a
nd
ne
xt
s
e
nt
e
nc
e
pr
e
di
c
ti
on
(
N
S
P
)
ta
s
ks
on
la
r
ge
-
s
c
a
le
c
or
por
a
to
e
xt
r
a
c
t
c
ont
e
xt
ua
l
s
e
m
a
nt
ic
in
f
or
m
a
ti
on
a
nd
obt
a
in
r
e
la
ti
ons
hi
ps
be
twe
e
n
s
e
nt
e
n
c
e
s
[
9]
–
[
11]
.
B
E
R
T
c
a
n
be
tt
e
r
e
xt
r
a
c
t
s
e
nt
e
nc
e
f
e
a
tu
r
e
in
f
or
m
a
ti
on
a
nd
unde
r
s
ta
nd
th
e
s
e
m
a
nt
ic
r
e
la
ti
ons
hi
ps
be
twe
e
n
s
e
nt
e
n
c
e
s
,
due
to
th
e
us
e
of
bi
di
r
e
c
ti
ona
l
e
nc
odi
ng
a
nd
s
e
lf
-
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
.
T
he
ti
tl
e
,
a
bs
tr
a
c
t,
a
nd
k
e
yw
or
ds
of
s
c
ie
nt
if
ic
pa
pe
r
s
a
r
e
th
e
m
os
t
im
por
ta
nt
m
e
ta
-
in
f
or
m
a
ti
on
of
th
e
pa
pe
r
,
in
c
lu
di
ng
s
e
m
a
nt
ic
in
f
or
m
a
ti
on
th
a
t
c
a
n
be
us
e
d
to
di
s
ti
ngui
s
h
di
f
f
e
r
e
nt
c
a
te
gor
ie
s
.
T
he
y c
a
n
be
e
a
s
il
y
obt
a
in
e
d
a
s
tr
a
in
in
g
c
or
pus
.
C
ur
r
e
nt
ly
,
th
e
r
e
is
li
tt
le
in
-
de
pt
h
r
e
s
e
a
r
c
h
on
th
e
r
ol
e
of
ti
tl
e
s
,
a
bs
tr
a
c
t
s
,
a
nd
ke
yw
or
ds
in
c
la
s
s
if
ic
a
ti
on
a
nd
how
th
e
y
a
r
e
us
e
d
in
c
om
bi
na
ti
on.
I
n
ou
r
s
tu
dy,
w
e
a
ppl
ie
d
th
e
pr
e
-
tr
a
in
e
d
la
ngua
ge
m
ode
l
B
E
R
T
a
nd
ot
he
r
de
e
p
le
a
r
ni
ng
m
ode
ls
s
uc
h
a
s
C
N
N
a
nd
L
S
T
M
f
or
pa
pe
r
c
la
s
s
if
ic
a
ti
on.
W
e
us
e
d
di
f
f
e
r
e
nt
c
om
bi
na
ti
ons
of
in
put
da
ta
f
e
a
tu
r
e
s
a
nd
m
ode
ls
,
a
nd
m
e
a
s
ur
e
d
pe
r
f
or
m
a
nc
e
th
r
ough
a
c
c
ur
a
c
y,
w
e
ig
ht
e
d
a
ve
r
a
ge
pr
e
c
is
io
n,
w
e
ig
ht
e
d
a
ve
r
a
ge
r
e
c
a
ll
, a
nd
w
e
ig
ht
e
d
a
ve
r
a
ge
F
1
-
s
c
or
e
. T
hi
s
p
a
pe
r
is
or
ga
ni
z
e
d
a
s
f
ol
lo
w
s
.
W
e
f
ir
s
t
i
nt
r
oduc
e
th
e
la
te
s
t
r
e
s
e
a
r
c
h
pr
ogr
e
s
s
in
p
a
pe
r
c
la
s
s
if
ic
a
ti
on,
th
e
n
in
tr
oduc
e
th
e
r
e
s
e
a
r
c
h
m
e
th
ods
in
c
lu
di
ng
da
ta
s
e
ts
a
nd
e
xpe
r
im
e
nt
a
l
s
e
tt
in
gs
,
th
e
n
di
s
c
us
s
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
t
s
,
a
nd
f
in
a
ll
y
c
onc
lu
de
.
2.
R
E
L
A
T
E
D
WORK
C
la
s
s
if
ic
a
ti
on
of
s
c
ie
nt
if
ic
pa
pe
r
s
us
in
g
m
a
c
hi
ne
l
e
a
r
ni
ng
m
e
th
ods
ha
s
be
e
n
e
xt
e
n
s
iv
e
ly
s
tu
di
e
d, a
nd
m
os
t
of
th
e
r
e
s
e
a
r
c
h
u
s
e
s
th
e
m
e
ta
da
ta
in
th
e
p
a
pe
r
s
,
th
a
t
is
,
to
e
xt
r
a
c
t
th
e
f
e
a
tu
r
e
ve
c
to
r
s
of
th
e
pa
pe
r
s
f
r
om
th
e
ti
tl
e
,
a
b
s
tr
a
c
t,
ke
yw
or
ds
,
a
nd
ot
he
r
i
nf
or
m
a
ti
on
to
tr
a
in
va
r
io
us
c
la
s
s
if
ie
r
s
.
S
e
v
e
r
a
l
r
e
s
e
a
r
c
he
r
s
[
12]
–
[
14]
pr
opos
e
d t
o us
e
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
c
la
s
s
if
ie
r
or
B
a
ye
s
ia
n
a
lg
or
it
hm
t
o r
e
a
li
z
e
t
he
c
la
s
s
if
ic
a
ti
on of
pa
pe
r
s
.
X
ia
ohua
a
nd H
a
iy
un
[
15]
pr
opos
e
d
a
hi
e
r
a
r
c
hi
c
a
l
c
la
s
s
if
ic
a
ti
on
m
e
th
od f
or
C
hi
ne
s
e
s
c
ie
nt
if
ic
pa
pe
r
s
ba
s
e
d
on
im
por
ta
nt
w
or
ds
i
n
t
it
le
s
, ke
yw
or
ds
, a
nd a
bs
tr
a
c
ts
. W
or
ds
i
n
th
e
pa
pe
r
t
e
xt
w
il
l
a
ls
o be
us
e
d i
f
t
he
y ha
ve
hi
gh
m
ut
ua
l
in
f
or
m
a
ti
on
va
lu
e
w
it
h
im
po
r
ta
nt
w
or
ds
.
F
or
w
o
r
ds
in
di
f
f
e
r
e
nt
pa
pe
r
a
r
e
a
s
,
a
β
va
lu
e
is
a
s
s
ig
ne
d
to
th
e
f
e
a
tu
r
e
ve
c
to
r
c
a
lc
ul
a
ti
on
f
o
r
m
ul
a
,
a
nd
th
e
va
lu
e
s
a
r
e
a
r
r
a
nge
d
in
th
e
f
ol
lo
w
in
g
or
de
r
ti
tl
e
>
s
um
m
a
r
y
>
ke
yw
or
ds
>
m
a
in
t
e
xt
.
U
s
in
g
de
e
p
le
a
r
ni
ng
te
c
hni
que
s
to
c
la
s
s
if
y
s
c
ie
nt
if
ic
pa
pe
r
s
ha
s
be
c
om
e
popula
r
in
r
e
c
e
nt
ye
a
r
s
.
C
h
ouy
ye
kh
e
t
al
.
[
16
]
pr
opo
s
e
d
to
us
e
C
N
N
t
o c
la
s
s
if
y
s
c
ie
nt
if
ic
pa
pe
r
s
, a
nd
us
e
d t
h
e
"
W
e
b of
S
c
ie
nc
e
D
a
ta
s
e
t"
a
s
a
n
e
xp
e
r
im
e
n
ta
l
da
t
a
s
e
t,
w
hi
c
h
c
ont
a
i
ns
i
np
ut
te
xt
s
e
q
ue
nc
e
,
t
a
r
g
e
t
la
be
l
va
lu
e
,
do
m
a
in
,
k
e
y
w
or
d
s
,
a
nd
s
u
m
m
a
r
y
in
f
or
m
a
t
io
n
of
35
23
8
pa
pe
r
s
.
B
ur
n
s
e
t
al
.
[
17
]
bui
lt
d
e
e
p
le
a
r
ni
n
g
m
o
de
ls
f
or
e
vi
d
e
n
c
e
c
l
a
s
s
if
i
c
a
ti
on
f
r
om
t
he
o
pe
n
-
a
c
c
e
s
s
bi
o
m
e
di
c
a
l
li
te
r
a
t
ur
e
,
de
ve
lo
p
e
d
a
l
a
r
g
e
-
s
c
a
le
c
or
pu
s
f
r
om
P
u
b
M
e
d
a
nd
P
u
bM
e
d
c
e
n
tr
a
l
ope
n
-
a
c
c
e
s
s
r
e
c
or
d
s
a
n
d
t
he
n
u
s
e
d
G
lo
ve
,
F
a
s
t
T
e
xt
, a
n
d
E
L
M
o a
lg
or
it
hm
s
t
o
l
e
a
r
n
w
or
d e
m
be
ddi
ng
. T
h
e
y a
l
s
o
us
e
C
N
N
, L
S
T
M
,
a
n
d a
tt
e
nt
io
n
m
e
c
h
a
n
i
s
m
s
t
o
im
pr
ov
e
t
he
e
f
f
e
c
t
of
c
la
s
s
if
ic
a
t
io
n
[
17]
. S
a
m
a
m
i
a
nd
S
our
e
[
18]
us
e
d
e
n
s
e
m
bl
e
d
e
e
p
le
a
r
ni
ng
m
ode
ls
to
c
la
s
s
if
y
L
upus
s
c
ie
nt
i
f
ic
a
r
ti
c
le
s
,
na
m
e
ly
,
a
c
om
bi
na
ti
on
of
L
S
T
M
,
c
uda
de
e
p
ne
ur
a
l
ne
twor
k
ga
te
d
r
e
c
ur
r
e
nt
uni
t
(
C
uD
N
N
G
R
U
)
,
R
N
N
,
a
nd
C
N
N
m
ode
ls
w
e
r
e
us
e
d
to
c
la
s
s
if
y
pa
pe
r
a
bs
tr
a
c
ts
,
a
nd
th
e
f
in
a
l
c
la
s
s
if
ic
a
ti
on
r
e
s
ul
ts
w
e
r
e
s
e
le
c
te
d
th
r
ough
vot
in
g,
a
nd
th
e
r
e
s
ul
ts
s
how
e
d
th
a
t
th
e
e
ns
e
m
bl
e
m
e
th
od
im
pr
ove
s
th
e
r
e
li
a
bi
li
ty
of
c
la
s
s
if
ic
a
ti
on
[
18]
.
B
ogda
nc
hi
kov
e
t
al
.
[
19]
us
e
d
a
de
e
p
le
a
r
ni
ng
m
ode
l
a
nd
na
iv
e
B
a
ye
s
a
lg
or
it
hm
to
c
l
a
s
s
if
y
s
c
ie
nt
if
ic
pa
pe
r
s
w
r
it
te
n
in
K
a
z
a
kh
la
ngua
g
e
,
a
nd
pr
oc
e
s
s
e
d
im
a
ge
a
nd
te
xt
s
e
pa
r
a
te
ly
.
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
t
s
s
h
ow
e
d
th
a
t
th
e
a
c
c
ur
a
c
y
w
a
s
im
pr
ove
d
by
us
in
g
m
ul
ti
m
oda
l
in
f
o
r
m
a
ti
on
c
om
pa
r
e
d
to
us
in
g
te
xt
f
e
a
tu
r
e
s
or
im
a
g
e
s
a
lo
ne
.
S
e
m
a
nt
ic
f
e
a
tu
r
e
d
c
onvolut
io
n
ne
ur
a
l
ne
twor
ks
(
S
F
-
C
N
N
)
w
e
r
e
p
r
opos
e
d
in
[
20]
to
im
p
r
ove
th
e
pe
r
f
or
m
a
nc
e
of
tr
a
di
t
io
na
l
C
N
N
w
hi
c
h
doe
s
not
c
ons
id
e
r
th
e
s
e
m
a
nt
ic
s
of
ba
g
-
of
-
w
or
ds
.
T
he
tr
a
in
in
g
da
ta
s
e
t
w
a
s
c
ol
le
c
te
d
f
r
om
A
r
X
iv
,
a
nd
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
e
d t
ha
t
th
e
c
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y r
e
a
c
he
d 94%
.
F
or
r
e
s
e
a
r
c
h
on
th
e
c
la
s
s
if
ic
a
ti
on
of
C
hi
ne
s
e
s
c
ie
nt
if
ic
pa
pe
r
s
,
L
il
i
e
t
al
.
[
21
]
us
e
d
th
e
B
E
R
T
m
ode
l
to
c
la
s
s
if
y
di
f
f
e
r
e
nt
ty
pe
s
of
C
hi
ne
s
e
li
te
r
a
tu
r
e
a
nd
a
c
hi
e
ve
d
a
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
of
76.95%
a
nd
68.55%
r
e
s
pe
c
ti
ve
ly
[
21]
.
A
not
he
r
s
tu
dy
a
ls
o
s
how
e
d
th
a
t
B
E
R
T
m
ode
ls
out
pe
r
f
or
m
e
d
th
e
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
m
ode
l,
a
m
ong
w
hi
c
h
th
e
B
E
R
T
-
re
-
pr
e
t
r
a
in
in
g
-
m
e
d
-
C
hi
ne
s
e
m
ode
l
pe
r
f
or
m
e
d
be
s
t
[
22
]
.
H
ongl
in
g
e
t
al
.
[
23
]
s
tu
di
e
d t
he
i
m
pa
c
t
of
s
to
p w
o
r
ds
i
n s
c
ie
nt
if
ic
pa
pe
r
s
on c
la
s
s
if
i
c
a
ti
on pe
r
f
or
m
a
nc
e
, a
nd t
he
y
f
ound tha
t
R
N
N
,
L
S
T
M
,
a
nd
ga
te
d
r
e
c
ur
r
e
nt
uni
t
(
G
R
U
)
m
ode
ls
c
oul
d
a
c
hi
e
ve
be
tt
e
r
pe
r
f
or
m
a
nc
e
w
it
hout
r
e
m
ovi
ng
s
to
p
w
or
ds
.
U
s
in
g
A
da
m
or
s
to
c
ha
s
ti
c
gr
a
di
e
nt
de
s
c
e
nt
(
S
G
D
)
opt
im
iz
e
r
f
or
R
N
N
a
nd
L
S
T
M
m
ode
ls
,
a
nd
A
d
a
de
lt
a
or
S
G
D
opt
im
iz
e
r
f
or
G
R
U
m
ode
ls
c
a
n
im
pr
ove
th
e
c
la
s
s
if
ic
a
ti
o
n
e
f
f
e
c
t
[
23]
.
J
ie
[
24]
de
ve
lo
pe
d
a
n
a
ut
om
a
ti
c
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
hi
ne
s
e
pape
r
c
la
s
s
if
ic
at
io
n bas
e
d on p
r
e
-
tr
ai
ne
d l
anguage
m
o
de
l
and hy
br
id
…
(
X
in
L
uo
)
643
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on
s
y
s
te
m
th
a
t
us
e
s
th
e
s
ki
m
-
gr
a
m
w
or
d
e
m
be
ddi
ng
m
ode
l
to
e
xt
r
a
c
t
th
e
f
e
a
tu
r
e
m
a
tr
ix
of
th
e
doc
um
e
nt
a
nd
a
dopt
s
th
e
C
N
N
m
ode
l
a
s
th
e
c
la
s
s
if
ie
r
.
T
he
f
ir
s
t
-
le
ve
l,
s
e
c
ond
-
le
ve
l,
a
nd
f
in
a
l
-
le
ve
l
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
w
e
r
e
97.66%
,
95.12%
, a
nd 92.42%
r
e
s
p
e
c
ti
ve
ly
[
24]
. Z
ha
ng
e
t
al
.
[
25]
s
tu
di
e
d t
he
r
ol
e
of
th
e
f
ul
l
-
te
xt
a
nd
s
tr
uc
tu
r
a
l
in
f
or
m
a
ti
on
of
pa
pe
r
s
in
c
la
s
s
if
ic
a
ti
on.
T
he
us
e
of
th
e
pr
e
-
tr
a
in
e
d
m
ode
l
L
ongF
or
m
e
r
s
how
e
d
th
a
t
th
e
in
tr
oduc
ti
on
of
f
ul
l
-
te
xt
in
f
or
m
a
ti
on
w
il
l
le
a
d
to
a
de
c
r
e
a
s
e
in
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y, w
hi
le
a
b
s
tr
a
c
t,
ke
yw
or
d, a
nd t
it
le
i
nf
or
m
a
ti
on pla
ys
a
de
c
is
iv
e
r
ol
e
i
n pa
pe
r
c
la
s
s
if
ic
a
ti
on
[
25]
.
3.
M
E
T
H
O
D
W
e
f
ir
s
t
dow
nl
oa
de
d
th
e
publ
ic
da
ta
C
hi
ne
s
e
s
c
ie
nt
if
ic
li
te
r
a
tu
r
e
(
C
S
L
)
da
ta
s
e
t
a
nd
pr
e
pr
oc
e
s
s
e
d
it
to
ge
ne
r
a
te
tr
a
in
in
g
da
ta
s
e
ts
,
de
ve
lo
pm
e
nt
d
a
ta
s
e
t
s
,
a
nd
te
s
t
d
a
ta
s
e
ts
.
T
h
e
n
w
e
de
s
ig
ne
d
f
our
ty
pe
s
of
e
xpe
r
im
e
nt
s
,
in
c
lu
di
ng
di
f
f
e
r
e
nt
c
om
bi
na
ti
ons
of
ke
yw
or
ds
,
ti
tl
e
s
,
a
nd
a
bs
tr
a
c
ts
in
s
in
gl
e
-
s
e
nt
e
nc
e
m
ode
,
s
in
gl
e
-
s
e
nt
e
nc
e
a
nd
s
e
nt
e
n
c
e
-
pa
ir
in
put
m
e
th
ods
,
e
n
s
e
m
bl
e
le
a
r
ni
ng
m
e
th
ods
,
a
nd
di
f
f
e
r
e
nt
m
ode
l
c
om
bi
na
ti
ons
,
to
c
om
pa
r
e
th
e
ir
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
unde
r
di
f
f
e
r
e
nt
pa
r
a
m
e
te
r
s
e
tt
in
gs
. C
la
s
s
if
ic
a
ti
on pe
r
f
or
m
a
nc
e
i
s
m
e
a
s
ur
e
d by
a
c
c
ur
a
c
y, pr
e
c
is
io
n, r
e
c
a
ll
, a
nd F
1 s
c
or
e
.
3.1.
T
h
e
d
at
as
e
t
A
la
r
ge
-
s
c
a
le
C
S
L
da
ta
s
e
t
w
a
s
bui
lt
in
[
26]
,
w
hi
c
h
c
ont
a
in
s
th
e
ti
tl
e
s
,
a
bs
tr
a
c
t
s
,
ke
yw
or
ds
,
a
nd
ot
he
r
f
ie
ld
s
of
396
k a
c
a
d
e
m
ic
pa
pe
r
s
,
i
t
s
houl
d be
t
h
e
f
ir
s
t
publ
ic
C
S
L
da
ta
s
e
t.
I
t
c
a
n b
e
us
e
d
f
or
N
L
P
ta
s
ks
s
u
c
h a
s
te
xt
s
um
m
a
r
iz
a
ti
on,
ke
yw
or
d
ge
ne
r
a
ti
on
,
a
nd
te
xt
c
la
s
s
if
ic
a
ti
on.
T
he
pa
pe
r
'
s
m
e
ta
-
in
f
or
m
a
ti
on
is
f
r
om
th
e
N
a
ti
ona
l
E
ngi
ne
e
r
in
g R
e
s
e
a
r
c
h C
e
nt
e
r
f
or
S
c
ie
nc
e
a
nd
T
e
c
hnol
ogy R
e
s
our
c
e
s
S
ha
r
in
g S
e
r
vi
c
e
(
N
S
T
R
)
, da
te
d
f
r
om
2010
to
2020.
T
he
c
ha
r
a
c
te
r
is
ti
c
s
of
th
e
da
ta
s
e
t
in
c
lu
d
e
w
id
e
r
di
s
c
ip
li
ne
c
ove
r
a
g
e
,
ne
w
d
a
ta
s
our
c
e
(
be
c
a
us
e
m
os
t
of
th
e
c
ur
r
e
nt
pa
pe
r
da
ta
s
e
ts
a
r
e
ta
ke
n
f
r
om
a
r
X
iv
,
P
ubM
e
d,
A
C
L
A
nt
hol
ogy
,
a
nd
M
A
G
)
,
hi
ghe
r
qua
li
ty
a
nd
a
c
c
ur
a
c
y
(
th
e
p
a
pe
r
ha
s
be
e
n
pe
e
r
-
r
e
vi
e
w
e
d)
.
T
he
s
ta
ti
s
ti
c
a
l
in
f
or
m
a
ti
on
of
th
e
e
nt
ir
e
da
ta
s
e
t
is
s
how
n
in
T
a
bl
e
1.
S
in
c
e
publ
ic
C
hi
ne
s
e
s
c
ie
nt
if
ic
p
a
pe
r
d
a
ta
s
e
ts
a
r
e
r
a
r
e
,
m
os
t
r
e
s
e
a
r
c
he
r
s
obt
a
in
th
e
m
e
ta
-
in
f
or
m
a
ti
on
of
pa
pe
r
s
f
r
om
th
e
li
br
a
r
y
li
te
r
a
tu
r
e
s
ha
r
in
g
pl
a
tf
or
m
or
onl
in
e
li
te
r
a
tu
r
e
da
ta
ba
s
e
s
,
a
nd
da
ta
s
e
ts
c
ons
tr
uc
te
d
by
th
e
m
a
r
e
ge
ne
r
a
ll
y
not
m
a
d
e
publ
ic
, s
o t
he
C
S
L
da
ta
s
e
t
c
a
n
pr
ovi
de
a
b
e
nc
hm
a
r
k
f
or
pe
r
f
or
m
a
nc
e
e
va
lu
a
ti
on of
C
hi
n
e
s
e
s
c
ie
nt
if
ic
p
a
pe
r
c
la
s
s
if
ic
a
ti
o
n m
ode
ls
. A
s
a
m
pl
e
f
r
om
t
hi
s
da
ta
s
e
t
i
s
s
ho
w
n
in
T
a
bl
e
2.
T
a
bl
e
1. D
e
ta
il
e
d
s
ta
ti
s
ti
c
s
of
t
he
C
S
L
da
ta
s
e
t
C
a
t
e
gor
y
#d
L
e
n(
T
)
L
e
n(
A
)
num
(
K
)
#s
a
m
pl
e
s
D
i
s
c
i
pl
i
ne
e
xa
m
pl
e
s
E
ngi
ne
e
r
i
ng
27
19.1
210.9
4.4
177,600
M
e
c
ha
ni
c
s
,
A
r
c
hi
t
e
c
t
ur
e
,
E
l
e
c
t
r
i
c
a
l
s
c
i
e
nc
e
S
c
i
e
nc
e
9
20.7
254.4
4.3
35,766
M
a
t
he
m
a
t
i
c
s
,
P
hy
s
i
c
s
,
A
s
t
r
onom
y, G
e
ogr
a
phy
A
gr
i
c
ul
t
ur
e
7
17.1
177.1
7.1
39,560
C
r
op
s
c
i
e
nc
e
, H
or
t
i
c
ul
t
ur
e
,
F
or
e
s
t
r
y
M
e
di
c
i
ne
5
20.7
269.5
4.7
36,783
C
l
i
ni
c
a
l
m
e
di
c
i
ne
, D
e
nt
a
l
m
e
di
c
i
ne
,
P
ha
r
m
a
c
y
M
a
na
ge
m
e
nt
4
18.7
157.7
6.2
23,630
B
us
i
ne
s
s
m
a
na
ge
m
e
nt
,
P
ubl
i
c
a
dm
i
ni
s
t
r
a
t
i
on
J
ur
i
s
pr
ude
nc
e
4
18.9
174.4
6.1
21,554
L
e
ga
l
s
c
i
e
nc
e
, P
ol
i
t
i
c
a
l
s
c
i
e
n
c
e
,
S
oc
i
ol
ogy
P
e
da
gogy
3
17.7
179.4
4.3
16,720
P
e
da
gogy, P
s
yc
hol
ogy,
P
hys
i
c
a
l
e
duc
a
t
i
on
E
c
onom
i
c
s
2
19.5
177.2
4.5
11,558
T
he
or
e
t
i
c
a
l
e
c
onom
i
c
s
,
A
ppl
i
e
d
e
c
onom
i
c
s
L
i
t
e
r
a
t
ur
e
2
18.8
158.2
8.3
10,501
C
hi
ne
s
e
l
i
t
e
r
a
t
ur
e
,
J
our
na
l
i
s
m
A
r
t
1
17.8
170.8
5.4
5,201
A
r
t
H
i
s
t
or
y
1
17.6
181.0
6.0
6,270
H
i
s
t
or
y
S
t
r
a
t
e
gi
c
s
1
17.5
169.3
4.0
3,555
M
i
l
i
t
a
r
y
s
c
i
e
nc
e
P
hi
l
os
ophy
1
18.0
176.5
8.0
7511
P
hi
l
os
ophy
A
l
l
67
396,209
#d:
t
he
num
be
r
of
di
s
c
i
pl
i
ne
s
i
n t
he
c
a
t
e
gor
y. l
e
n(
T
)
:
a
ve
r
a
ge
l
e
ngt
h of
e
a
c
h
t
i
t
l
e
;
l
e
n(
A
)
:
a
ve
r
a
ge
l
e
ngt
h of
e
a
c
h a
bs
t
r
a
c
t
;
num
(
K
)
:
a
ve
r
a
ge
num
be
r
of
ke
yw
or
ds
T
a
bl
e
2. A
s
a
m
pl
e
f
r
om
t
he
C
S
L
da
ta
s
e
t
in
E
ngl
is
h
T
i
t
l
e
A
bs
t
r
a
c
t
K
e
yw
or
ds
D
i
s
c
i
pl
i
ne
C
a
t
e
gor
y
E
xpl
or
a
t
i
on
on
i
m
pr
ovi
ng
pe
a
s
a
nt
s
’
s
c
i
e
nt
i
f
i
c
a
nd
c
ul
t
ur
a
l
qua
l
i
t
y
by
us
i
ng
di
s
t
a
nc
e
e
duc
a
t
i
on
S
t
a
r
t
i
ng f
r
o
m
t
he
i
m
por
t
a
nc
e
of
i
m
p
r
ovi
ng
f
a
r
m
e
r
s
’
s
c
i
e
nt
i
f
i
c
a
nd
t
e
c
hnol
ogi
c
a
l
qua
l
i
t
y,
t
hi
s
pa
pe
r
di
s
c
us
s
e
s
t
he
w
a
y
s
a
nd
c
ha
r
a
c
t
e
r
i
s
t
i
c
s
of
di
s
t
a
nc
e
e
duc
a
t
i
on
P
e
da
gogy
;
D
i
s
t
a
nc
e
e
duc
a
t
i
on
;
C
om
m
e
nt
s
;
Q
ua
l
i
t
y
A
gr
i
c
ul
t
ur
a
l
E
ngi
ne
e
r
i
ng
E
ngi
ne
e
r
i
ng
3.2.
F
e
at
u
r
e
e
xt
r
ac
t
io
n
B
E
R
T
is
a
n ope
n
-
s
our
c
e
m
a
c
hi
ne
l
e
a
r
ni
ng f
r
a
m
e
w
or
k f
or
N
L
P
.
B
E
R
T
i
s
de
s
ig
n
e
d t
o he
lp
c
om
put
e
r
s
unde
r
s
ta
nd
th
e
m
e
a
ni
ng
of
a
m
bi
guous
la
ngua
ge
in
th
e
te
xt
by
us
in
g
s
ur
r
ounding
te
xt
to
e
s
ta
bl
is
h
c
ont
e
xt
.
I
n
m
a
ny
na
tu
r
a
l
la
ngua
ge
unde
r
s
ta
ndi
ng
(
N
L
U
)
ta
s
ks
,
s
uc
h
a
s
s
e
n
ti
m
e
nt
a
na
ly
s
is
,
s
e
m
a
nt
ic
r
ol
e
a
nnot
a
ti
on,
a
nd
te
xt
c
la
s
s
if
ic
a
ti
on,
B
E
R
T
c
a
n
a
c
hi
e
ve
be
tt
e
r
pe
r
f
or
m
a
nc
e
th
a
n
ot
he
r
de
e
p
le
a
r
ni
ng
m
e
th
ods
.
T
he
pr
oc
e
s
s
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
20
25
:
641
-
649
644
us
in
g
B
E
R
T
to
ge
n
e
r
a
te
th
e
to
ke
n
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on
ve
c
to
r
of
th
e
te
xt
s
e
que
nc
e
is
a
s
f
ol
lo
w
.
F
or
E
ngl
is
h
te
xt
s
e
que
n
c
e
s
,
w
or
d
s
e
gm
e
nt
a
ti
on
i
s
f
ir
s
t
pe
r
f
or
m
e
d,
w
hi
le
f
or
C
hi
ne
s
e
te
xt
s
e
que
nc
e
s
,
it
is
pr
oc
e
s
s
e
d
c
ha
r
a
c
te
r
by
c
ha
r
a
c
te
r
,
a
nd
f
in
a
ll
y
f
or
m
s
a
s
e
que
nc
e
c
om
pos
e
d
of
E
ngl
is
h
w
or
ds
or
C
hi
ne
s
e
c
ha
r
a
c
t
e
r
to
ke
ns
.
[
C
L
S
]
a
nd
[
S
E
P
]
ta
gs
a
r
e
us
e
d
to
r
e
pr
e
s
e
nt
th
e
c
la
s
s
if
ic
a
ti
on
a
nd
s
e
p
a
r
a
to
r
of
s
e
nt
e
nc
e
s
a
nd
w
il
l
b
e
a
dde
d
to
th
e
be
gi
nni
ng
a
nd
e
nd
of
th
e
s
e
nt
e
nc
e
r
e
s
pe
c
ti
ve
ly
.
T
he
e
m
be
ddi
ng
r
e
pr
e
s
e
nt
a
ti
on
of
e
a
c
h
to
ke
n
is
obt
a
in
e
d
f
r
om
voc
a
b,
a
voc
a
bul
a
r
y pr
ovi
de
d by
B
E
R
T
.
T
he
s
um
of
t
ok
e
n e
m
be
ddi
ng,
s
e
gm
e
nt
e
m
be
ddi
ng,
a
nd pos
it
io
n
e
m
be
ddi
ng i
s
in
put
to
th
e
tr
a
ns
f
or
m
e
r
la
ye
r
,
a
nd
th
e
to
ke
n
ve
c
to
r
w
it
h
gl
oba
l
s
e
m
a
nt
ic
in
f
or
m
a
ti
on
a
nd
c
la
s
s
ve
c
to
r
w
il
l
be
ge
ne
r
a
te
d.
T
he
pr
oc
e
s
s
is
s
how
n
in
F
ig
ur
e
1,
c
la
s
s
ve
c
to
r
a
nd
to
ke
n
ve
c
to
r
a
r
e
de
not
e
d
by
[
]
a
nd
r
e
s
pe
c
ti
ve
ly
.
T
he
[
]
ve
c
to
r
or
a
ve
r
a
ge
va
lu
e
of
ot
he
r
to
ke
n
ve
c
t
or
s
out
put
by
B
E
R
T
w
il
l
be
in
put
to
th
e
c
la
s
s
if
ic
a
ti
on l
a
ye
r
t
o c
om
pl
e
te
t
he
p
a
pe
r
c
la
s
s
if
ic
a
ti
on.
F
ig
ur
e
1. P
a
pe
r
c
la
s
s
if
ic
a
ti
on by us
in
g B
E
R
T
m
ode
l
3.3.
C
la
s
s
if
i
c
at
io
n
m
od
e
li
n
g
C
N
N
a
nd
R
N
N
m
ode
ls
r
e
qui
r
e
pr
e
pr
oc
e
s
s
in
g
th
e
te
xt
s
e
qu
e
nc
e
s
be
f
or
e
c
la
s
s
if
ic
a
ti
on.
A
f
te
r
th
e
te
xt
s
e
que
nc
e
s
w
e
r
e
t
oke
ni
z
e
d, a
voc
a
bul
a
r
y w
a
s
ge
ne
r
a
te
d. W
or
d e
m
be
ddi
ng ve
c
to
r
s
c
a
n be
r
a
ndoml
y ge
ne
r
a
te
d,
but
us
in
g publi
c
pr
e
-
tr
a
in
e
d w
or
d e
m
be
ddi
ng
ve
c
to
r
s
c
a
n
a
c
hi
e
ve
be
tt
e
r
c
l
a
s
s
if
ic
a
ti
on
r
e
s
ul
ts
.
I
n our
s
t
udy,
t
he
C
hi
ne
s
e
pr
e
-
tr
a
in
e
d
w
or
d
ve
c
to
r
s
w
e
r
e
us
e
d,
w
hi
c
h
w
e
r
e
obt
a
in
e
d
by
tr
a
in
in
g
th
e
W
or
d2ve
c
m
ode
l
on
th
e
B
a
id
u E
nc
yc
lo
pe
di
a
c
or
pus
(
W
or
d+
C
ha
r
a
c
te
r
, 300d)
[
27]
.
3.4.
T
h
e
e
xp
e
r
im
e
n
t
s
e
t
u
p
T
o
tr
a
in
a
nd
te
s
t
th
e
m
ode
ls
,
w
e
e
xt
r
a
c
te
d
1,000
r
e
c
or
ds
pe
r
di
s
c
ip
li
ne
f
r
om
th
e
C
S
L
da
ta
s
e
t
to
c
ons
tr
uc
t
th
e
e
xpe
r
im
e
nt
a
l
da
ta
s
e
t,
in
c
lu
di
ng
a
bs
tr
a
c
t,
k
e
yw
or
ds
,
ti
tl
e
,
a
nd
c
a
te
gor
y
f
ie
ld
s
.
T
he
e
xpe
r
im
e
nt
a
l
da
ta
s
e
t
i
s
di
vi
de
d
in
to
a
tr
a
in
in
g
s
e
t,
va
li
da
ti
o
n
s
e
t,
a
nd
te
s
t
s
e
t
a
c
c
or
di
ng
to
th
e
r
a
ti
o
of
7:
2:
1.
T
h
e
ha
r
dw
a
r
e
c
onf
ig
ur
a
ti
on
of
th
e
e
xpe
r
im
e
nt
a
l
pl
a
tf
or
m
a
nd
m
ode
l
de
ve
lo
p
m
e
nt
f
r
a
m
e
w
or
k
a
r
e
s
how
n
in
T
a
bl
e
3.
W
e
us
e
d
f
our
m
e
th
ods
t
o
c
om
pa
r
e
a
nd a
na
ly
z
e
t
he
pe
r
f
or
m
a
nc
e
of
pa
pe
r
c
la
s
s
if
ic
a
ti
on, i
nc
lu
di
n
g s
in
gl
e
s
e
nt
e
nc
e
i
nput
a
nd c
onc
a
te
na
ti
on me
th
od,
s
e
nt
e
nc
e
p
a
ir
i
nput
m
e
th
od, e
ns
e
m
b
le
l
e
a
r
ni
ng me
th
od, a
nd c
om
bi
na
ti
on of
B
E
R
T
a
nd
C
N
N
/R
N
N
m
ode
ls
.
T
he
de
f
in
it
io
ns
of
th
e
s
e
m
e
th
ods
a
nd
th
e
e
xpe
r
im
e
nt
a
l
pur
pos
e
s
a
r
e
s
how
n
in
T
a
bl
e
4.
C
las
s
if
ie
r
E
[
C
L
S
]
E
1
E
n
E
[
S
E
P
]
……
A
bs
t
r
a
c
t
/
K
e
yw
or
ds
/
T
i
t
l
e
[
C
L
S
]
T
ok
1
T
ok
n
[
S
E
P
]
……
T
r
a
ns
f
or
m
e
r
1
T
r
a
ns
f
or
m
e
r
n
T
[
C
L
S
]
T
1
T
n
T
[
S
E
P
]
…
…
B
E
R
T
……
F
ul
l
C
onne
c
t
e
d L
a
ye
r
+ S
of
t
m
a
x
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
hi
ne
s
e
pape
r
c
la
s
s
if
ic
at
io
n bas
e
d on p
r
e
-
tr
ai
ne
d l
anguage
m
o
de
l
and hy
br
id
…
(
X
in
L
uo
)
645
T
a
bl
e
3. E
xpe
r
im
e
nt
a
l
e
nvi
r
onm
e
nt
pa
r
a
m
e
te
r
s
P
a
r
a
m
e
t
e
r
V
a
l
ue
C
P
U
I
nt
e
l
(
R
)
X
e
on(
R
)
S
i
l
ve
r
4216,2.8G
R
A
M
128G
G
P
U
R
T
X
3090
M
ode
l
de
ve
l
opm
e
nt
f
r
a
m
e
w
or
k
P
yT
or
c
h 1.13.1
T
a
bl
e
4. D
if
f
e
r
e
nt
m
e
th
ods
f
or
pa
pe
r
c
la
s
s
if
ic
a
ti
on
N
o
.
M
e
t
hod
D
e
s
c
r
i
pt
i
on
P
ur
pos
e
1
S
i
ngl
e
s
e
nt
e
nc
e
i
nput
a
nd
c
onc
a
t
e
na
t
i
on
m
e
t
hod
A
bs
t
r
a
c
t
,
ke
yw
or
ds
,
a
nd
t
i
t
l
e
a
r
e
r
e
ga
r
de
d
a
s
a
s
i
ngl
e
s
e
nt
e
nc
e
a
nd
i
nput
i
nt
o
t
he
B
E
R
T
m
ode
l
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on.
U
s
e
di
f
f
e
r
e
nt
m
e
t
hods
t
o
c
onc
a
t
e
na
t
e
a
b
s
t
r
a
c
t
, ke
yw
or
ds
, a
nd t
i
t
l
e
be
f
or
e
e
nt
e
r
i
ng t
he
m
ode
l
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on.
C
om
pa
r
e
t
he
i
m
por
t
a
nc
e
of
t
he
A
bs
t
r
a
c
t
,
ke
yw
or
ds
,
a
nd
t
i
t
l
e
i
n
pa
pe
r
c
l
a
s
s
i
f
i
c
a
t
i
on, a
nd a
n
a
l
yz
e
t
he
i
m
pa
c
t
of
c
onne
c
t
i
ng
m
ul
t
i
pl
e
f
i
e
l
ds
a
s
i
nput
t
o
t
he
m
ode
l
on c
l
a
s
s
i
f
i
c
a
t
i
on pe
r
f
or
m
a
nc
e
.
2
S
e
nt
e
nc
e
pa
i
r
s
i
nput
m
e
t
hod
C
om
bi
ne
t
he
A
bs
t
r
a
c
t
,
ke
yw
or
ds
,
a
nd
t
i
t
l
e
i
nt
o
s
e
nt
e
nc
e
pa
i
r
s
i
n
di
f
f
e
r
e
nt
w
a
ys
,
a
nd
t
he
n
i
nput
t
he
m
i
nt
o t
he
B
E
R
T
m
ode
l
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on.
A
na
l
yz
e
t
he
i
m
pa
c
t
of
di
f
f
e
r
e
nt
c
om
bi
na
t
i
on
m
e
t
hods
on
c
l
a
s
s
i
f
i
c
a
t
i
on
pe
r
f
or
m
a
nc
e
i
n t
he
s
e
nt
e
nc
e
pa
i
r
m
ode
.
3
E
ns
e
m
bl
e
l
e
a
r
ni
ng
m
e
t
hod
T
he
B
E
R
T
m
ode
l
i
s
u
s
e
d
t
o
c
l
a
s
s
i
f
y
A
b
s
t
r
a
c
t
s
,
ke
yw
or
ds
,
a
nd
t
i
t
l
e
s
r
e
s
pe
c
t
i
ve
l
y,
a
nd
t
he
c
l
a
s
s
i
f
i
c
a
t
i
on
r
e
s
ul
t
s
of
t
he
t
hr
e
e
m
ode
l
s
a
r
e
i
nt
e
gr
a
t
e
d
t
o
obt
a
i
n
t
he
f
i
na
l
c
l
a
s
s
i
f
i
c
a
t
i
on
r
e
s
ul
t
.
C
om
pa
r
e
t
he
i
m
pa
c
t
of
i
nt
e
gr
a
t
i
ng
t
he
c
l
a
s
s
i
f
i
c
a
t
i
on r
e
s
ul
t
s
of
t
he
t
hr
e
e
m
ode
l
s
i
n
di
f
f
e
r
e
nt
w
a
ys
on
t
he
f
i
na
l
c
l
a
s
s
i
f
i
c
a
t
i
on r
e
s
ul
t
s
.
4
C
om
bi
na
t
i
on
of
B
E
R
T
a
nd
C
N
N
/
L
S
T
M
m
ode
l
U
s
e
t
he
B
E
R
T
m
ode
l
t
o
obt
a
i
n
t
he
s
e
nt
e
nc
e
r
e
pr
e
s
e
nt
a
t
i
on
f
e
a
t
ur
e
ve
c
t
or
o
f
t
he
a
bs
t
r
a
c
t
,
a
nd
t
he
n
i
nput
i
t
i
nt
o
t
he
C
N
N
/
L
S
T
M
m
ode
l
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on.
C
om
pa
r
e
t
he
pa
pe
r
c
l
a
s
s
i
f
i
c
a
t
i
on
pe
r
f
or
m
a
nc
e
w
he
n
c
om
bi
ni
ng
B
E
R
T
w
i
t
h ot
he
r
m
ode
l
s
.
I
n
or
de
r
to
e
va
lu
a
te
th
e
pe
r
f
or
m
a
nc
e
of
th
e
m
ode
l
in
pa
pe
r
c
la
s
s
if
ic
a
ti
on,
w
e
in
tr
oduc
e
4
m
e
tr
ic
s
,
in
c
lu
di
ng
a
c
c
ur
a
c
y,
w
e
ig
ht
e
d
a
ve
r
a
ge
pr
e
c
is
io
n,
w
e
ig
ht
e
d
a
v
e
r
a
ge
r
e
c
a
ll
,
a
nd
w
e
ig
ht
e
d
a
ve
r
a
ge
F
1
-
s
c
or
e
,
w
hi
c
h a
r
e
de
f
in
e
d i
n (
1)
to
(
4)
.
Ac
c
ur
a
c
y
=
TP
+
TN
TP
+
FP
+
TN
+
FN
(
1)
P
r
e
c
is
io
n
=
TP
TP
+
FP
(
2)
R
e
c
a
l
l
=
TP
TP
+
FN
(
3)
F
1
−
s
c
o
r
e
=
2TP
2TP
+
FP
+
FN
(
4)
A
m
ong
th
e
m
,
tr
ue
pos
it
iv
e
(
T
P
)
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
c
or
r
e
c
tl
y
pr
e
di
c
te
d
pos
it
iv
e
s
a
m
pl
e
s
,
f
a
ls
e
pos
it
iv
e
(
F
P
)
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
in
c
or
r
e
c
tl
y
pr
e
di
c
te
d
pos
it
iv
e
s
a
m
pl
e
s
,
tr
ue
ne
g
a
ti
ve
(
T
N
)
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
c
or
r
e
c
tl
y
pr
e
di
c
te
d
ne
ga
ti
ve
s
a
m
pl
e
s
a
nd
f
a
ls
e
n
e
ga
ti
ve
(
F
N
)
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
ne
ga
ti
ve
s
a
m
pl
e
s
in
c
or
r
e
c
tl
y
pr
e
di
c
te
d.
S
in
c
e
p
a
pe
r
c
la
s
s
if
ic
a
ti
on
i
s
a
m
ul
ti
-
c
la
s
s
if
ic
a
ti
on
pr
obl
e
m
,
w
e
c
a
n
c
a
lc
ul
a
te
th
e
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
va
lu
e
s
of
e
a
c
h
c
a
te
gor
y
f
ir
s
t,
a
nd
th
e
n
th
e
w
e
ig
ht
e
d
a
ve
r
a
g
e
pr
e
c
is
io
n,
w
e
ig
ht
e
d
a
ve
r
a
ge
r
e
c
a
ll
,
a
nd
w
e
ig
ht
e
d
a
ve
r
a
ge
F
1
-
s
c
or
e
m
e
tr
ic
s
a
r
e
us
e
d
to
m
e
a
s
ur
e
th
e
ove
r
a
ll
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
f
or
a
ll
c
a
te
gor
ie
s
.
T
a
ki
ng
th
e
w
e
ig
ht
e
d
a
ve
r
a
ge
F1
-
s
c
or
e
a
s
a
n
e
xa
m
pl
e
,
it
s
c
a
lc
ul
a
ti
on
f
or
m
ul
a
is
in
(
5)
.
W
e
ig
ht
e
d
Ave
r
a
g
e
F
1
−
s
c
o
r
e
=
F
1
cl
a
s
s
1
W
1
+
F
1
cl
a
s
s
2
W
2
+
.
.
.
+
F
1
cl
a
s
s
N
W
N
(
5)
F
1
cl
a
s
s
N
r
e
pr
e
s
e
nt
s
F
1
-
s
c
or
e
of
c
la
s
s
N
,
W
N
r
e
pr
e
s
e
nt
s
th
e
w
e
ig
ht
of
e
a
c
h
c
la
s
s
,
w
hi
c
h
is
de
t
e
r
m
in
e
d
by
th
e
r
a
ti
o of
t
he
numbe
r
of
s
a
m
pl
e
s
i
n c
la
s
s
N
t
o t
he
t
ot
a
l
num
be
r
of
s
a
m
pl
e
s
.
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
tu
dy
in
ve
s
ti
ga
te
s
th
e
di
f
f
e
r
e
nt
e
f
f
e
c
ts
of
ti
tl
e
s
,
ke
y
w
or
ds
,
a
nd
a
bs
tr
a
c
t
f
ie
ld
s
on
pa
pe
r
c
la
s
s
if
ic
a
ti
on. Although ea
r
li
e
r
s
tu
di
e
s
e
xpl
or
e
d t
he
e
f
f
e
c
ts
of
i
ndi
vi
dua
l
f
ie
ld
s
, t
he
y di
d not e
xpl
ic
it
ly
a
ddr
e
s
s
th
e
e
f
f
e
c
ts
of
di
f
f
e
r
e
nt
f
ie
ld
c
om
bi
na
ti
ons
.
W
e
h
a
ve
te
s
te
d
th
r
e
e
di
f
f
e
r
e
nt
c
om
bi
na
ti
on
m
e
th
ods
in
B
E
R
T
m
ode
l.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
20
25
:
641
-
649
646
4
.1.
S
in
gl
e
s
e
n
t
e
n
c
e
i
n
p
u
t
an
d
c
on
c
at
e
n
at
io
n
m
e
t
h
od
T
he
in
put
of
te
xt
s
e
que
nc
e
f
or
th
e
B
E
R
T
m
ode
l
c
a
n
be
in
th
e
f
or
m
of
a
s
in
gl
e
s
e
nt
e
nc
e
or
a
pa
ir
of
s
e
nt
e
nc
e
s
.
T
o
e
va
lu
a
te
th
e
im
pa
c
t
of
th
e
a
bs
tr
a
c
t,
ke
yw
or
ds
,
a
n
d
ti
tl
e
of
th
e
pa
pe
r
on
th
e
c
la
s
s
if
ic
a
ti
on
e
f
f
e
c
t
of
th
e
pa
pe
r
,
th
e
a
bs
tr
a
c
t,
ke
yw
or
ds
,
a
nd
ti
tl
e
a
r
e
tr
e
a
te
d
a
s
in
de
pe
nd
e
nt
s
e
nt
e
nc
e
s
,
a
nd
th
e
n
u
s
e
th
e
B
E
R
T
/C
N
N
/L
S
T
M
m
od
e
l
to
c
a
lc
ul
a
te
th
e
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
in
th
e
e
xpe
r
im
e
nt
a
l
da
ta
s
e
t.
I
n
a
ddi
ti
on,
di
f
f
e
r
e
nt
c
onc
a
te
na
te
d
f
or
m
s
be
twe
e
n
a
bs
tr
a
c
ts
,
ke
yw
or
ds
,
a
nd
ti
tl
e
s
,
in
c
lu
di
ng
a
bs
tr
a
c
t+
ti
tl
e
,
a
bs
tr
a
c
t+
ke
yw
or
d
s
,
a
nd
a
bs
tr
a
c
t+
k
e
yw
or
ds
+
ti
tl
e
,
a
r
e
tr
e
a
te
d
a
s
s
e
nt
e
nc
e
s
to
e
va
lu
a
te
th
e
im
pa
c
t
of
di
f
f
e
r
e
nt
ty
pe
s
of
in
f
or
m
a
ti
on
c
om
bi
na
ti
ons
on
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
.
T
he
e
xp
e
r
im
e
nt
a
l
r
e
s
ul
ts
a
r
e
s
how
n
in
T
a
bl
e
5.
I
t
c
a
n
be
s
e
e
n
th
a
t
us
in
g
a
bs
tr
a
c
t,
ke
yw
or
ds
,
a
nd
ti
tl
e
a
lo
ne
a
s
th
e
in
put
da
ta
of
th
e
B
E
R
T
m
ode
l,
in
put
ti
ng
a
bs
tr
a
c
t
c
a
n a
c
hi
e
ve
t
he
hi
ghe
s
t
c
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y, w
hi
c
h i
s
s
ig
ni
f
ic
a
nt
ly
hi
ghe
r
t
ha
n ke
yw
or
ds
or
ti
tl
e
.
C
om
pa
r
e
d
w
it
h
C
N
N
a
nd
L
S
T
M
m
ode
ls
,
th
e
B
E
R
T
m
ode
l
c
a
n
a
c
hi
e
ve
hi
gh
e
r
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
.
F
or
s
e
nt
e
nc
e
s
c
ont
a
in
in
g
m
or
e
th
a
n
two
e
le
m
e
nt
s
f
r
om
a
bs
tr
a
c
t,
ke
yw
or
ds
,
a
nd
ti
tl
e
,
th
e
m
e
th
ods
th
a
t
c
om
bi
ni
ng a
bs
tr
a
c
ts
w
it
h ke
yw
or
ds
,
or
c
om
bi
ni
ng a
bs
tr
a
c
t
s
w
it
h t
it
le
s
, c
a
n
s
li
ght
ly
i
m
pr
ove
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
ove
r
us
i
ng
onl
y
one
e
le
m
e
nt
.
H
ow
e
ve
r
,
th
e
c
la
s
s
if
ic
a
ti
on
e
f
f
e
c
t
of
c
om
bi
ni
ng
a
bs
tr
a
c
t,
ke
yw
or
ds
a
nd
ti
tl
e
s
is
not
a
s
good
a
s
th
e
f
or
m
e
r
.
I
t
s
how
s
th
a
t
w
he
n
m
o
r
e
s
e
nt
e
nc
e
in
f
or
m
a
ti
on
is
in
put
, m
or
e
noi
s
e
da
ta
m
a
y
a
ls
o be
i
m
por
te
d.
T
a
bl
e
5
. S
in
gl
e
s
e
nt
e
nc
e
&
c
onc
a
te
na
ti
on me
th
od c
la
s
s
if
ic
a
ti
on
M
ode
l
I
nput
da
t
a
(
S
i
ngl
e
s
e
nt
e
nc
e
)
A
c
c
ur
a
c
y
W
e
i
ght
e
d
a
ve
r
a
ge
pr
e
c
i
s
i
on
W
e
i
ght
e
d
a
ve
r
a
ge
r
e
c
a
l
l
W
e
i
ght
e
d
a
ve
r
a
ge
F
1 s
c
or
e
B
E
R
T
A
bs
t
r
a
c
t
0.8690
0.8689
0.8688
0.8673
C
N
N
A
bs
t
r
a
c
t
0.8007
0.7975
0.8007
0.7970
L
S
T
M
A
bs
t
r
a
c
t
0.7939
0.7908
0.7939
0.7899
B
E
R
T
T
i
t
l
e
0.8214
0.8181
0.8214
0.8182
B
E
R
T
K
e
yw
or
ds
0.8200
0.8200
0.8201
0.8185
B
E
R
T
A
bs
t
r
a
c
t
+
title
0.8720
0.8723
0.8725
0.8709
B
E
R
T
A
bs
t
r
a
c
t
+
ke
yw
or
ds
0.8707
0.8701
0.8707
0.8693
B
E
R
T
A
bs
t
r
a
c
t
+ke
yw
or
ds
+
title
0.8680
0.8679
0.8681
0.8664
4.2. S
e
n
t
e
n
c
e
p
ai
r
s
i
n
p
u
t
m
e
t
h
od
I
nput
ti
ng
th
e
a
bs
tr
a
c
t,
k
e
yw
or
ds
a
nd
ti
tl
e
of
th
e
pa
pe
r
in
to
th
e
B
E
R
T
m
ode
l
in
th
e
f
or
m
of
s
e
nt
e
nc
e
pa
ir
s
a
c
tu
a
ll
y a
ll
ow
s
t
he
m
ode
l
to
l
e
a
r
n t
he
r
e
la
ti
on
s
hi
p be
twe
e
n t
he
t
w
o s
e
nt
e
n
c
e
s
.
T
he
c
a
te
gor
y t
o w
hi
c
h t
he
pa
pe
r
be
lo
ngs
c
a
n
be
r
e
ga
r
de
d
a
s
a
r
e
la
ti
ons
hi
p.
C
la
s
s
if
ic
a
ti
on
of
pa
pe
r
s
is
a
c
hi
e
ve
d
by
le
a
r
ni
ng
th
e
im
pl
ic
it
a
s
s
oc
ia
ti
on
in
f
or
m
a
ti
on
of
A
bs
tr
a
c
t,
ke
yw
or
ds
a
nd
ti
tl
e
.
T
he
t
hr
e
e
ty
pe
s
of
s
e
nt
e
nc
e
pa
ir
s
<
A
bs
tr
a
c
t,
T
it
le
>
,
<
A
bs
tr
a
c
t,
K
e
y
w
or
ds
>
,
<
A
bs
tr
a
c
t,
T
it
le
+
ke
y
w
or
ds
>
w
il
l
be
us
e
d
a
s
in
put
to
th
e
B
E
R
T
m
ode
l,
a
nd
th
e
ir
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
w
il
l
be
e
va
lu
a
te
d.
A
m
ong
th
e
m
,
<
A
bs
tr
a
c
t,
T
it
le
+
ke
y
w
or
ds
>
m
e
a
n
s
th
a
t
ti
tl
e
a
nd
ke
yw
or
ds
a
r
e
f
ir
s
t
c
onc
a
te
na
te
d
in
to
a
s
e
nt
e
nc
e
,
a
nd
th
e
n
c
o
m
bi
ne
d
w
it
h
a
bs
tr
a
c
t
to
f
or
m
a
s
e
nt
e
nc
e
pa
ir
.
A
c
c
or
di
ng
to
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
,
th
e
c
la
s
s
if
ic
a
ti
on
e
f
f
e
c
t
of
s
e
nt
e
nc
e
pa
ir
s
us
in
g
th
e
<
A
bs
tr
a
c
t,
T
it
le
+
ke
yw
or
d
s
>
m
e
th
od
in
th
e
B
E
R
T
m
ode
l
is
s
li
ght
ly
be
tt
e
r
th
a
n
th
e
ot
he
r
two
m
e
th
ods
.
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
r
e
s
how
n i
n
T
a
bl
e
6.
T
a
bl
e
6
. S
e
nt
e
nc
e
pa
ir
c
la
s
s
if
ic
a
ti
on r
e
s
ul
t
M
ode
l
I
nput
D
a
t
a
(
S
e
nt
e
nc
e
P
a
i
r
)
A
c
c
ur
a
c
y
W
e
i
ght
e
d
a
ve
r
a
ge
pr
e
c
i
s
i
on
W
e
i
ght
e
d
a
ve
r
a
ge
r
e
c
a
l
l
W
e
i
ght
e
d
a
ve
r
a
ge
F
1 s
c
or
e
S
e
nt
e
nc
e
A
S
e
nt
e
nc
e
B
B
E
R
T
A
bs
t
r
a
c
t
T
i
t
l
e
0.8811
0.8806
0.8811
0.8788
B
E
R
T
A
bs
t
r
a
c
t
K
e
yw
or
ds
0.8860
0.8846
0.8858
0.8841
B
E
R
T
A
bs
t
r
a
c
t
T
i
t
l
e
+ K
e
yw
or
ds
0.8880
0.8867
0.8875
0.8865
4.3. E
n
s
e
m
b
le
l
e
a
r
n
in
g m
e
t
h
od
I
t
c
a
n
be
s
e
e
n
f
r
om
th
e
pr
e
vi
ous
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
th
a
t
th
e
B
E
R
T
m
ode
l
is
us
e
d
to
c
la
s
s
if
y
th
e
a
bs
tr
a
c
t,
ke
yw
or
ds
a
nd t
it
le
of
t
he
pa
p
e
r
, a
nd t
he
a
c
c
ur
a
c
y i
s
86
.9%
, 82.0%
a
nd 82.1%
r
e
s
pe
c
ti
ve
ly
. A
lt
hough
th
e
s
e
que
nc
e
le
ngt
h
of
a
bs
tr
a
c
ts
f
a
r
e
xc
e
e
d
s
ke
yw
or
ds
a
nd
ti
tl
e
s
,
th
e
la
tt
e
r
two
ty
pe
s
of
te
xt
s
e
que
nc
e
s
s
ti
ll
c
ont
a
in
im
por
ta
nt
in
f
or
m
a
ti
on
th
a
t
c
a
n
di
s
ti
ngui
s
h
di
f
f
e
r
e
nt
c
a
te
gor
ie
s
.
T
he
r
e
f
or
e
,
in
te
gr
a
ti
ng
th
e
out
put
r
e
s
ul
ts
of
th
e
B
E
R
T
m
ode
ls
a
f
te
r
pr
oc
e
s
s
in
g
th
e
a
bs
tr
a
c
t,
ke
yw
or
ds
a
n
d
ti
tl
e
r
e
s
pe
c
ti
ve
ly
m
a
y
im
pr
ove
th
e
f
in
a
l
pa
pe
r
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
T
h
e
[
C
L
S
]
to
ke
n
out
put
v
e
c
to
r
s
obt
a
i
ne
d
a
f
te
r
th
e
B
E
R
T
m
ode
l
pr
oc
e
s
s
e
s
a
bs
tr
a
c
t,
ke
yw
or
ds
,
a
nd
ti
tl
e
r
e
s
pe
c
ti
ve
ly
a
r
e
r
e
pr
e
s
e
nt
e
d
a
s
B
E
R
T
_A
bs
tr
a
c
t,
B
E
R
T
_K
e
yw
or
d
a
nd
B
E
R
T
_
T
it
le
in
tu
r
n.
T
he
[
C
S
L
]
out
put
ve
c
to
r
s
of
di
f
f
e
r
e
nt
ty
pe
s
a
r
e
s
um
m
e
d
a
nd
th
e
n
in
put
in
to
th
e
c
la
s
s
if
ic
a
ti
on
la
ye
r
to
c
om
pl
e
t
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
hi
ne
s
e
pape
r
c
la
s
s
if
ic
at
io
n bas
e
d on p
r
e
-
tr
ai
ne
d l
anguage
m
o
de
l
and hy
br
id
…
(
X
in
L
uo
)
647
th
e
c
la
s
s
if
ic
a
ti
on
of
th
e
pa
pe
r
.
E
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
th
a
t
th
e
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
of
s
um
m
in
g
B
E
R
T
_A
bs
tr
a
c
t,
B
E
R
T
_K
e
yw
or
d
a
nd
B
E
R
T
_T
it
le
a
s
th
e
in
put
of
th
e
c
la
s
s
if
ie
r
is
be
tt
e
r
th
a
n
us
in
g
ot
he
r
m
e
th
ods
,
s
uc
h
a
s
s
um
m
in
g
B
E
R
T
_A
bs
tr
a
c
t
a
nd
B
E
R
T
_K
e
yw
or
d,
s
um
m
in
g
B
E
R
T
_A
b
s
tr
a
c
t
a
nd
B
E
R
T
_
T
it
le
,
or
us
in
g B
E
R
T
_A
bs
tr
a
c
t
a
lo
ne
. T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
r
e
s
h
ow
n i
n T
a
bl
e
7.
T
a
bl
e
7.
E
ns
e
m
bl
e
le
a
r
ni
ng me
th
od c
la
s
s
if
ic
a
ti
on r
e
s
ul
t
M
ode
l
I
nput
da
t
a
(
S
i
ngl
e
s
e
nt
e
nc
e
)
A
c
c
ur
a
c
y
W
e
i
ght
e
d
a
ve
r
a
ge
pr
e
c
i
s
i
on
W
e
i
ght
e
d
a
ve
r
a
ge
r
e
c
a
l
l
W
e
i
ght
e
d
a
ve
r
a
ge
F
1 s
c
or
e
B
E
R
T
B
E
R
T
_A
bs
t
r
a
c
t
0.8690
0.8689
0.8688
0.8673
B
E
R
T
B
E
R
T
_A
bs
t
r
a
c
t
+B
E
R
T
_K
e
yw
or
ds
0.8730
0.8766
0.8730
0.8724
B
E
R
T
B
E
R
T
_A
bs
t
r
a
c
t
+B
E
R
T
_T
i
t
l
e
0.8689
0.8694
0.8689
0.8684
B
E
R
T
B
E
R
T
_A
bs
t
r
a
c
t
+B
E
R
T
_K
e
yw
or
ds
+B
E
R
T
_T
i
t
l
e
0.8740
0.8743
0.8745
0.8719
4.4. Com
b
in
at
io
n
of
B
E
R
T
an
d
C
N
N
/L
S
T
M
m
od
e
l
s
S
in
c
e
th
e
B
E
R
T
m
od
e
l
is
good
a
t
a
c
qui
r
in
g
s
e
m
a
nt
ic
in
f
or
m
a
ti
on
of
te
xt
s
e
que
nc
e
s
,
th
e
te
xt
f
e
a
tu
r
e
ve
c
to
r
s
it
out
put
s
c
a
n
b
e
us
e
d
a
s
in
put
f
e
a
tu
r
e
ve
c
to
r
s
f
or
ot
he
r
m
ode
ls
,
s
o
th
a
t
th
e
a
dva
nt
a
ge
s
of
va
r
io
us
m
ode
ls
c
a
n
be
c
om
pr
e
he
ns
iv
e
ly
ut
il
iz
e
d
to
im
pr
ove
th
e
pe
r
f
or
m
a
nc
e
of
pa
pe
r
c
la
s
s
if
ic
a
ti
on.
W
e
u
s
e
d
th
e
pa
pe
r
a
b
s
tr
a
c
t
a
s
in
put
da
ta
,
obt
a
in
ed
th
e
to
ke
n'
s
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on
ve
c
to
r
th
r
ough
th
e
B
E
R
T
m
ode
l,
a
nd
th
e
n
in
put
it
in
to
th
e
C
N
N
,
L
S
T
M
,
a
nd
R
C
N
N
m
ode
ls
f
or
c
la
s
s
if
ic
a
ti
on.
T
h
e
C
N
N
m
ode
l
u
s
e
d
256
c
onvolut
io
n
ke
r
ne
l
s
w
it
h
s
iz
e
s
of
1,
2
a
nd
3,
m
a
xi
m
um
pool
in
g
m
e
th
od
is
us
e
d
to
r
e
duc
e
th
e
di
m
e
ns
io
n
of
th
e
out
put
f
e
a
tu
r
e
s
.
T
he
R
C
N
N
m
ode
l
w
il
l
us
e
t
he
f
or
m
ul
a
in
(
6)
.
O
ut
p
ut
=
M
a
x
P
o
o
l
(
L
S
T
M
(
B
ER
T
_
o
ut
p
ut
)
+
B
ER
T
_
o
ut
p
ut
)
(
6)
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
r
e
s
ho
w
n
in
T
a
bl
e
8
.
I
t
c
a
n
be
s
e
e
n
th
a
t
th
e
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
of
th
e
B
E
R
T
+
C
N
N
a
nd
B
E
R
T
+
R
C
N
N
m
ode
l
s
i
s
i
m
pr
ove
d c
om
pa
r
e
d t
o t
he
B
E
R
T
m
ode
l
a
lo
n
e
.
T
hi
s
s
how
s
t
ha
t
th
e
C
N
N
m
ode
l'
s
a
bi
li
ty
to
obt
a
in
lo
c
a
l
f
e
a
tu
r
e
s
of
t
e
xt
s
e
que
nc
e
s
c
a
n
im
pr
ove
th
e
c
la
s
s
if
ic
a
ti
on
p
e
r
f
or
m
a
nc
e
of
t
he
B
E
R
T
m
ode
l.
T
a
bl
e
8.
C
om
bi
na
ti
on of
B
E
R
T
a
nd othe
r
m
ode
ls
M
ode
l
I
nput
da
t
a
(
S
i
ngl
e
s
e
nt
e
nc
e
)
A
c
c
ur
a
c
y
W
e
i
ght
e
d
a
ve
r
a
ge
pr
e
c
i
s
i
o
n
W
e
i
ght
e
d
a
ve
r
a
ge
r
e
c
a
l
l
W
e
i
ght
e
d
a
ve
r
a
ge
F
1 s
c
or
e
B
E
R
T
A
bs
t
r
a
c
t
0.8690
0.8689
0.8688
0.8673
B
E
R
T
+C
N
N
A
bs
t
r
a
c
t
0.8739
0.8736
0.8739
0.8725
B
E
R
T
+L
S
T
M
A
bs
t
r
a
c
t
0.8681
0.8700
0.8681
0.8683
B
E
R
T
+R
C
N
N
A
bs
t
r
a
c
t
0.8737
0.8743
0.8737
0.8721
5.
C
O
N
C
L
U
S
I
O
N
I
n
or
de
r
to
s
tu
dy
how
to
e
f
f
e
c
ti
ve
ly
ut
il
iz
e
a
bs
tr
a
c
t,
ke
yw
o
r
d
,
a
nd
ti
tl
e
in
f
or
m
a
ti
on
to
a
c
hi
e
ve
a
ut
om
a
ti
c
c
la
s
s
if
ic
a
ti
on
of
C
hi
ne
s
e
p
a
pe
r
s
,
m
ul
ti
pl
e
in
put
da
ta
p
r
oc
e
s
s
in
g
m
e
th
ods
,
a
nd
m
ul
ti
pl
e
de
e
p
le
a
r
ni
ng
m
ode
ls
w
e
r
e
a
ppl
ie
d t
o
t
he
e
xp
e
r
im
e
nt
a
l
da
ta
s
e
t.
F
in
a
ll
y, w
e
c
a
n dr
a
w
t
he
f
ol
lo
w
in
g
c
onc
lu
s
io
ns
:
i
)
t
he
e
f
f
e
c
t
of
us
in
g
th
e
B
E
R
T
m
od
e
l
a
lo
ne
to
c
la
s
s
if
y
pa
p
e
r
s
is
s
ig
ni
f
ic
a
nt
ly
be
tt
e
r
th
a
n
us
in
g
th
e
C
N
N
or
L
S
T
M
m
ode
l
a
lo
ne
;
ii
)
in
pa
pe
r
c
la
s
s
if
ic
a
ti
on,
us
in
g
a
bs
tr
a
c
t
a
lo
n
e
a
s
th
e
in
p
ut
da
ta
of
th
e
B
E
R
T
m
ode
l,
th
e
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y of
t
he
m
ode
l
is
s
i
gni
f
ic
a
nt
ly
be
tt
e
r
t
ha
n us
in
g k
e
yw
or
ds
or
t
it
le
a
lo
ne
a
s
i
nput
da
ta
. A
f
te
r
c
onne
c
ti
ng
a
bs
tr
a
c
t,
ti
tl
e
a
nd
ke
yw
or
ds
in
di
f
f
e
r
e
nt
w
a
ys
a
s
in
put
da
ta
,
t
he
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
is
not
s
ig
ni
f
ic
a
nt
ly
im
pr
ove
d;
ii
i)
c
om
bi
ne
a
bs
tr
a
c
t,
ti
tl
e
,
a
nd
k
e
yw
or
ds
in
di
f
f
e
r
e
nt
w
a
ys
in
to
s
e
nt
e
nc
e
pa
ir
s
a
s
in
put
d
a
ta
f
or
th
e
B
E
R
T
m
ode
l.
T
h
e
c
la
s
s
if
ic
a
ti
on
p
e
r
f
or
m
a
nc
e
is
s
ig
ni
f
ic
a
nt
ly
i
m
pr
ove
d
c
om
pa
r
e
d
to
th
e
s
in
gl
e
s
e
nt
e
nc
e
in
put
f
or
m
;
iv
)
th
e
B
E
R
T
m
ode
l
is
us
e
d
to
p
r
oc
e
s
s
a
bs
tr
a
c
t,
ti
tl
e
,
a
nd
ke
yw
or
ds
r
e
s
pe
c
ti
ve
ly
,
a
nd
th
e
th
r
e
e
ou
tp
ut
r
e
s
ul
ts
a
r
e
s
um
m
e
d
in
di
f
f
e
r
e
nt
c
om
bi
na
ti
ons
a
nd
u
s
e
d
a
s
in
put
da
ta
f
or
th
e
c
la
s
s
if
ic
a
ti
on
la
ye
r
.
T
h
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
t
ha
t
th
e
c
la
s
s
if
ic
a
ti
on e
f
f
e
c
t
is
s
li
ght
ly
be
tt
e
r
by t
a
ki
ng t
he
s
um
of
t
he
out
put
r
e
s
ul
ts
of
th
e
th
r
e
e
B
E
R
T
m
ode
ls
th
a
n
us
i
ng
one
B
E
R
T
m
ode
l
a
lo
n
e
;
v)
tr
e
a
t
th
e
B
E
R
T
m
ode
l
a
s
a
te
xt
f
e
a
tu
r
e
e
xt
r
a
c
to
r
,
a
nd
th
e
obt
a
in
e
d
te
xt
f
e
a
tu
r
e
ve
c
to
r
s
a
r
e
th
e
n
in
put
in
to
C
N
N
,
R
N
N
,
or
ot
he
r
m
ode
ls
f
or
s
e
c
onda
r
y
pr
oc
e
s
s
in
g,
to
c
om
pr
e
he
ns
iv
e
ly
a
ppl
y
th
e
c
a
pa
bi
li
ti
e
s
of
di
f
f
e
r
e
nt
ty
pe
s
of
de
e
p
le
a
r
ni
ng
m
ode
ls
a
nd
e
xt
r
a
c
t
m
or
e
e
f
f
e
c
ti
ve
c
la
s
s
if
ic
a
ti
on i
nf
or
m
a
ti
on.
E
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
t
ha
t
th
e
c
om
bi
na
ti
on of
B
E
R
T
a
nd C
N
N
c
a
n
e
na
bl
e
th
e
m
ode
l
to
a
c
hi
e
ve
be
tt
e
r
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
th
a
n
ot
he
r
c
om
bi
na
ti
on
m
e
th
ods
.
H
ow
e
ve
r
,
c
om
pa
r
e
d
to
us
in
g
th
e
B
E
R
T
m
ode
l
a
lo
ne
f
or
c
la
s
s
if
ic
a
ti
on,
th
e
pe
r
f
or
m
a
nc
e
im
pr
ove
m
e
nt
is
not
obvi
ous
;
a
nd
vi
)
th
e
e
xpe
r
im
e
nt
a
l
d
a
ta
s
e
t
it
s
e
lf
a
l
s
o
ha
s
f
a
c
to
r
s
th
a
t
a
f
f
e
c
t
c
la
s
s
if
ic
a
ti
on
p
e
r
f
or
m
a
nc
e
.
C
ur
r
e
nt
ly
,
tr
a
in
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
20
25
:
641
-
649
648
s
a
m
pl
e
s
a
r
e
a
s
s
ig
ne
d
onl
y on
e
c
a
t
e
gor
y va
lu
e
.
S
om
e
pa
pe
r
s
be
l
ong to i
nt
e
r
di
s
c
ip
li
na
r
y a
nd
t
he
r
e
w
il
l
be
s
c
ope
ove
r
la
p
be
twe
e
n
c
a
te
gor
ie
s
.
T
he
r
e
f
or
e
,
hi
e
r
a
r
c
hi
c
a
l
a
nd
w
e
ig
ht
e
d
m
ul
ti
-
c
a
te
gor
y
pa
pe
r
c
l
a
s
s
if
ic
a
ti
on
is
m
or
e
pr
om
is
in
g.
I
n
th
e
ne
xt
s
te
p
of
r
e
s
e
a
r
c
h,
w
e
w
il
l
a
ls
o
tr
y
to
u
s
e
th
e
c
it
a
ti
on
in
f
or
m
a
ti
on
of
th
e
p
a
pe
r
a
s
a
s
uppl
e
m
e
nt
a
r
y f
ie
ld
t
o i
m
pr
ove
t
he
e
f
f
e
c
t
of
pa
pe
r
c
la
s
s
if
ic
a
ti
o
n.
R
E
F
E
R
E
N
C
E
S
[
1]
Q
.
L
i
e
t
al
.
,
“
A
s
ur
ve
y
on
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on:
f
r
om
t
r
a
di
t
i
ona
l
t
o
de
e
p
l
e
a
r
ni
n
g,”
A
C
M
T
r
ans
ac
t
i
on
s
on
I
nt
e
l
l
i
ge
nt
Sy
s
t
e
m
s
and
T
e
c
hnol
ogy
, vol
. 13, no. 2, 2022, doi
:
10.1145/
3495162.
[
2]
A
.
G
a
s
pa
r
e
t
t
o,
M
.
M
a
r
c
uz
z
o,
A
.
Z
a
nga
r
i
,
a
nd
A
.
A
l
ba
r
e
l
l
i
,
“
A
s
ur
ve
y
on
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on
a
l
gor
i
t
hm
s
:
f
r
om
t
e
xt
t
o
pr
e
di
c
t
i
o
n
s
,”
I
nf
or
m
at
i
on
, vol
. 13, no. 2, 2022, doi
:
10.3390/
i
nf
o13020083.
[
3]
T
.
Y
ue
,
Y
.
L
i
,
X
.
S
hi
,
J
.
Q
i
n,
Z
.
F
a
n,
a
nd
Z
.
H
u,
“
P
a
pe
r
N
e
t
:
A
da
t
a
s
e
t
a
nd
be
nc
h
m
a
r
k
f
or
f
i
ne
-
g
r
a
i
ne
d
pa
pe
r
c
l
a
s
s
i
f
i
c
a
t
i
on,”
A
ppl
i
e
d
Sc
i
e
nc
e
s
, vol
. 12, no. 9, 2022, doi
:
10.3390/
a
pp12094554.
[
4]
A
.
R
a
j
a
n
a
nd
M
.
M
a
nur
,
“
A
s
p
e
c
t
ba
s
e
d
s
e
nt
i
m
e
nt
a
na
l
ys
i
s
us
i
ng
f
i
ne
-
t
une
d
B
E
R
T
m
ode
l
w
i
t
h
de
e
p
c
ont
e
xt
f
e
a
t
ur
e
s
,”
I
A
E
S
I
nt
e
r
nat
i
onal
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
, vol
. 13, no. 2, pp. 1250
–
1261, 2024, doi
:
10.11591/
i
j
a
i
.v13.i
2.pp1250
-
1261.
[
5]
M
.
G
a
o,
T
.
L
i
,
a
nd
P
.
H
ua
ng,
“
T
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on
r
e
s
e
a
r
c
h
ba
s
e
d
on
i
m
pr
ove
d
w
or
d2ve
c
a
nd
C
N
N
,”
i
n
Se
r
v
i
c
e
-
O
r
i
e
nt
e
d
C
om
put
i
ng
–
I
C
SO
C
2018 W
or
k
s
hop
s
, 2019, pp. 126
–
135, doi
:
10.1007/
978
-
3
-
030
-
17642
-
6_11.
[
6]
S
.
A
bdul
-
R
a
hm
a
n,
M
.
F.
A.
M
.
A
l
i
,
A
.
A
.
B
a
ka
r
,
a
nd
S
.
M
ut
a
l
i
b
,
“
E
nha
nc
i
ng
c
hur
n
f
or
e
c
a
s
t
i
ng
w
i
t
h
s
e
nt
i
m
e
nt
a
na
l
y
s
i
s
of
s
t
e
a
m
r
e
vi
e
w
s
,”
Soc
i
al
N
e
t
w
o
r
k
A
nal
y
s
i
s
and M
i
ni
ng
, vol
.
14,
no. 178, pp. 1
–
17, 2024,
doi
:
10.1007/
s
13278
-
024
-
01337
-
3
[
7]
S
.
L
a
i
,
L
.
X
u,
K
.
L
i
u,
a
nd
J
.
Z
ha
o,
“
R
e
c
ur
r
e
nt
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
f
or
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on,”
P
r
oc
e
e
di
ng
s
of
t
he
N
at
i
onal
C
onf
e
r
e
nc
e
on A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
, vol
. 3, pp. 2267
–
2273, 2015, doi
:
10.1609/
a
a
a
i
.v29i
1.9513.
[
8]
C
. Z
hou, C
. S
un, Z
. L
i
u, a
nd F
. C
. M
.
L
a
u,
“
A
C
-
L
S
T
M
N
e
ur
a
l
N
e
t
w
or
k f
or
T
e
xt
C
l
a
s
s
i
f
i
c
a
t
i
on,”
a
r
X
i
v
-
C
om
put
e
r
Sc
i
e
n
c
e
, pp. 1
–
10, 2015.
[
9]
J
.
D
e
vl
i
n,
M
.
-
W
.
C
ha
ng,
K
.
L
e
e
,
K
.
T
.
G
oogl
e
,
a
nd
A
.
I
.
L
a
ngua
g
e
,
“
B
E
R
T
:
P
r
e
-
t
r
a
i
ni
ng
of
de
e
p
bi
di
r
e
c
t
i
ona
l
t
r
a
ns
f
or
m
e
r
s
f
or
l
a
ngua
ge
unde
r
s
t
a
ndi
ng,”
i
n
P
r
oc
e
e
di
ngs
of
N
A
A
C
L
-
H
L
T
2019
, 2019, pp. 4171
–
4186.
[
10]
I
.
B
e
l
t
a
gy,
K
.
L
o,
a
nd
A
.
C
oha
n,
“
S
C
I
B
E
R
T
:
A
pr
e
t
r
a
i
ne
d
l
a
ngua
ge
m
ode
l
f
o
r
s
c
i
e
nt
i
f
i
c
t
e
xt
,”
i
n
E
M
N
L
P
-
I
J
C
N
L
P
2019
-
2019
C
onf
e
r
e
nc
e
on
E
m
pi
r
i
c
al
M
e
t
hods
i
n
N
at
ur
al
L
anguage
P
r
oc
e
s
s
i
ng
and
9t
h
I
n
t
e
r
nat
i
onal
J
oi
nt
C
onf
e
r
e
nc
e
on
N
at
ur
al
L
anguag
e
P
r
oc
e
s
s
i
ng, P
r
oc
e
e
di
ngs
of
t
he
C
onf
e
r
e
nc
e
, 2019, pp. 3615
–
3620, doi
:
10.18653/
v1/
d19
-
1371.
[
11]
M
a
r
ya
nt
o,
P
hi
l
i
ps
,
a
nd
A
.
S
.
G
i
r
s
a
ng,
“
H
ybr
i
d
m
ode
l
f
o
r
e
xt
r
a
c
t
i
ve
s
i
ngl
e
doc
um
e
nt
s
um
m
a
r
i
z
a
t
i
on:
ut
i
l
i
z
i
ng
B
E
R
T
opi
c
a
nd
B
E
R
T
m
ode
l
,”
I
A
E
S
I
nt
e
r
nat
i
onal
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
,
vol
.
1
3,
no.
2,
pp.
1723
–
1731,
J
un.
2024,
doi
:
10.11591/
i
j
a
i
.v13.i
2.pp1723
-
1731.
[
12]
D
.
W
e
i
a
nd
Z
.
J
i
e
,
“
S
c
i
e
nt
i
f
i
c
l
i
t
e
r
a
t
ur
e
c
l
a
s
s
i
f
i
c
a
t
i
on
r
e
s
e
a
r
c
h
b
a
s
e
d
on
t
he
de
ns
i
t
y
di
s
t
r
i
but
i
on
of
O
C
S
V
M
,”
I
nf
or
m
at
i
on
E
ngi
ne
e
r
i
ng
, vol
. 4, no. 3, pp. 67
–
72, 2018, doi
:
10.3772/
j
.i
s
s
n.2095
-
915x.2018.03.009.
[
13]
M
.
I
.
E
l
i
a
s
,
Y
.
M
a
hm
ud,
S
.
M
ut
a
l
i
b,
S
.
N
.
K
.
K
a
m
a
r
udi
n,
R
.
M
a
s
ka
t
a
nd
S
.
A
.
R
a
hm
a
n,
“
F
a
ke
ne
w
s
pr
e
di
c
t
i
on
u
s
i
ng
hybr
i
d
m
ode
l
–
s
ys
t
e
m
a
t
i
c
l
i
t
e
r
a
t
ur
e
r
e
vi
e
w
,”
2023
4t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
A
r
t
i
f
i
c
i
a
l
I
nt
e
l
l
i
ge
nc
e
and
D
at
a
Sc
i
e
nc
e
s
(
A
i
D
A
S)
,
I
P
O
H
,
M
a
l
a
ys
i
a
, pp. 281
-
286, 2023, doi
:
10.1109/
A
i
D
A
S
60501.2023.10284628.
[
14]
I
.
J
a
ya
,
I
.
A
ul
i
a
,
S
.
M
.
H
a
r
di
,
J
.
T
.
T
a
r
i
ga
n,
M
.
S
.
L
ydi
a
,
a
nd
C
a
r
ol
i
ne
,
“
S
c
i
e
nt
i
f
i
c
doc
um
e
nt
s
c
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
s
uppor
t
ve
c
t
or
m
a
c
hi
ne
a
l
gor
i
t
hm
,”
J
our
nal
of
P
hy
s
i
c
s
:
C
onf
e
r
e
nc
e
Se
r
i
e
s
,
vol
.
1235,
no.
1,
pp.
1
-
6,
M
a
y
2019,
doi
:
10.1088/
1742
-
6596/
1235/
1/
012082.
[
15]
Y
.
X
i
a
ohua
a
nd
G
.
H
a
i
yun,
“
I
m
pr
ove
d
B
a
ye
s
i
a
n
a
l
gor
i
t
hm
ba
s
e
d
a
ut
om
a
t
i
c
c
l
a
s
s
i
f
i
c
a
t
i
on
m
e
t
hod
f
or
bi
bl
i
ogr
a
phy,”
C
om
put
e
r
Sc
i
e
nc
e
, vol
. 45, no. 8, pp. 203
–
207, 2018, doi
:
10.11896/
j
.i
s
s
n.1002
-
137X
.201
8.08.036.
[
16]
M
.
E
c
h
-
C
houyye
kh,
H
.
O
m
a
r
a
,
a
nd
M
.
L
a
z
a
a
r
,
“
S
c
i
e
nt
i
f
i
c
pa
pe
r
c
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
,”
i
n
A
C
M
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
P
r
o
c
e
e
di
ng Se
r
i
e
s
, 2019, pp. 1
–
6, doi
:
10.1145/
3372
938.3372951.
[
17]
G
.
A
.
B
ur
n
s
,
X
.
L
i
,
a
nd
N
.
P
e
ng,
“
B
ui
l
di
ng
de
e
p
l
e
a
r
ni
ng
m
ode
l
s
f
or
e
vi
de
n
c
e
c
l
a
s
s
i
f
i
c
a
t
i
on
f
r
om
t
he
op
e
n
a
c
c
e
s
s
bi
om
e
di
c
a
l
l
i
t
e
r
a
t
ur
e
,”
D
at
abas
e
, vol
. 2019, no. 1, 2019, doi
:
10.1093/
da
t
a
ba
s
e
/
ba
z
034.
[
18]
M
.
S
a
m
a
m
i
a
nd
E
.
M
.
S
our
e
,
“
B
i
na
r
y
c
l
a
s
s
i
f
i
c
a
t
i
on
of
L
upu
s
s
c
i
e
nt
i
f
i
c
a
r
t
i
c
l
e
s
a
ppl
yi
ng
d
e
e
p
e
n
s
e
m
bl
e
m
od
e
l
on
t
e
xt
da
t
a
,”
i
n
2019
Se
v
e
nt
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
D
i
gi
t
al
I
nf
or
m
at
i
on
P
r
oc
e
s
s
i
ng
and
C
om
m
uni
c
at
i
ons
(
I
C
D
I
P
C
)
,
2019,
pp.
12
–
17,
doi
:
10.1109/
I
C
D
I
P
C
.2019.8723787.
[
19]
A
.
B
ogda
nc
hi
kov,
D
.
A
y
a
z
ba
ye
v,
a
nd
I
.
V
a
r
l
a
m
i
s
,
“
C
l
a
s
s
i
f
i
c
a
t
i
on
of
s
c
i
e
nt
i
f
i
c
doc
um
e
nt
s
i
n
t
he
ka
z
a
kh
l
a
ngua
ge
u
s
i
ng
de
e
p
ne
u
r
a
l
ne
t
w
or
ks
a
nd
a
f
us
i
on
of
i
m
a
ge
s
a
nd
t
e
xt
,”
B
i
g
D
at
a
and
C
ogni
t
i
v
e
C
om
put
i
ng
,
vol
.
6,
no.
4,
pp.
1
-
12,
O
c
t
.
2022,
doi
:
10.3390/
bdc
c
6040123.
[
20]
R
.
S
a
r
a
s
u,
K
.
K
.
T
hya
gha
r
a
j
a
n,
a
nd
N
.
R
.
S
ha
nke
r
,
“
S
F
-
C
N
N
:
D
e
e
p
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on
a
nd
r
e
t
r
i
e
va
l
f
or
t
e
xt
doc
um
e
nt
s
,”
I
nt
e
l
l
i
ge
nt
A
ut
om
at
i
on and Sof
t
C
om
put
i
ng
, vol
. 35, no. 2, pp. 1799
–
1813, 2023, doi
:
10.3
2604/
i
a
s
c
.2023.027429.
[
21]
S
.
L
i
l
i
,
J
.
P
e
ng,
a
nd
W
.
J
i
ng,
“
A
s
t
udy
on
t
he
a
ut
om
a
t
i
c
c
l
a
s
s
i
f
i
c
a
t
i
on
of
c
hi
ne
s
e
l
i
t
e
r
a
t
ur
e
i
n
pe
r
i
odi
c
a
l
s
ba
s
e
d
on
B
E
R
T
m
o
de
l
,”
L
i
br
ar
y
J
our
nal
, vol
. 41, no. 5, 2022, doi
:
10.13663/
j
.c
nki
.l
j
.2022.05.014.
[
22]
Z
.
Y
a
ng,
Z
.
Z
hi
xi
ong,
L
.
H
ua
n,
a
nd
D
.
L
i
a
ngpi
ng,
“
C
l
a
s
s
i
f
i
c
a
t
i
on
of
c
hi
ne
s
e
m
e
di
c
a
l
l
i
t
e
r
a
t
ur
e
w
i
t
h
be
r
t
m
ode
l
,”
D
at
a
A
nal
y
s
i
s
and
K
now
l
e
dge
D
i
s
c
ov
e
r
y
, vol
. 4, no. 8, pp. 41
–
49, 2020, doi
:
10.11925/
i
nf
ot
e
c
h.2096
-
3467.2019.1238.
[
23]
X
.
H
ongl
i
ng,
F
.
G
uoh
e
,
a
nd
H
.
W
e
i
l
i
n,
“
R
e
s
e
a
r
c
h
on
s
e
m
a
nt
i
c
c
l
a
s
s
i
f
i
c
a
t
i
on
of
s
c
i
e
nt
i
f
i
c
a
nd
t
e
c
hni
c
a
l
l
i
t
e
r
a
t
ur
e
ba
s
e
d
on
de
e
p
l
e
a
r
ni
ng,”
I
nf
or
m
at
i
on
s
t
udi
e
s
:
T
he
or
y
&
A
ppl
i
c
at
i
on
,
vol
.
41,
no.
11,
pp.
149
–
154,
2018,
doi
:
10.16353/
j
.c
nki
.1000
-
7490.2018.11.027.
[
24]
K
.
J
i
e
,
“
R
e
s
e
a
r
c
h
on
a
ut
om
a
t
i
c
l
i
t
e
r
a
t
ur
e
c
l
a
s
s
i
f
i
c
a
t
i
on
s
ys
t
e
m
ba
s
e
d
on
de
e
p
l
e
a
r
ni
ng
a
nd
c
hi
ne
s
e
l
i
br
a
r
y
c
l
a
s
s
i
f
i
c
a
t
i
on,”
N
e
w
C
e
nt
ur
y
L
i
br
a
r
y
, vol
. 5, pp. 51
–
56, 2021, doi
:
10.16810/
j
.c
nki
.1672
-
514X
.2021.05.009.
[
25]
Y
.
Z
ha
ng
e
t
al
.
,
“
W
e
a
kl
y
s
upe
r
vi
s
e
d
m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on
of
f
ul
l
-
t
e
xt
s
c
i
e
nt
i
f
i
c
pa
pe
r
s
,”
i
n
K
D
D
'
23:
P
r
oc
e
e
di
ngs
of
t
he
29t
h
A
C
M
SI
G
K
D
D
C
onf
e
r
e
nc
e
on
K
now
l
e
dge
D
i
s
c
o
v
e
r
y
and
D
at
a
M
i
ni
ng
,
L
ong
B
e
a
c
h,
C
a
l
i
f
or
ni
a
,
A
ug.
2023,
pp.
3458
–
3469,
doi
:
10.1145/
3580305.3599544.
[
26]
Y
.
L
i
e
t
al
.
,
“
C
S
L
:
A
l
a
r
ge
-
s
c
a
l
e
c
hi
ne
s
e
s
c
i
e
nt
i
f
i
c
l
i
t
e
r
a
t
ur
e
da
t
a
s
e
t
,”
P
r
oc
e
e
d
i
ngs
-
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
, C
O
L
I
N
G
,
vol
. 29, no. 1, pp. 3917
–
3923, 2022.
[
27]
S
.
L
i
,
Z
.
Z
ha
o,
R
.
H
u,
W
.
L
i
,
T
.
L
i
u,
a
nd
X
.
D
u,
“
A
na
l
ogi
c
a
l
r
e
a
s
oni
ng
on
c
hi
ne
s
e
m
or
phol
ogi
c
a
l
a
nd
s
e
m
a
nt
i
c
r
e
l
a
t
i
ons
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
56t
h A
nnual
M
e
e
t
i
ng of
t
he
A
s
s
oc
i
at
i
on f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
,
M
e
l
bour
ne
, A
u
s
t
r
a
l
i
a
, vol
. 2, J
ul
.
2018,
pp. 138
–
143, doi
:
10.18653/
v1/
P
18
-
2023.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
hi
ne
s
e
pape
r
c
la
s
s
if
ic
at
io
n bas
e
d on p
r
e
-
tr
ai
ne
d l
anguage
m
o
de
l
and hy
br
id
…
(
X
in
L
uo
)
649
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Xin
Luo
is
pursuing
in
Computer
Scienc
e
in
School
of
Compu
ting
Sciences,
College
of
Computing,
Informa
tics
and
Mathema
tics,
Univers
iti
Teknologi
MARA
,
Shah
Alam,
Selangor,
Malaysia.
His
current
research
interests
are
d
eep
learning
and
natural
language
process.
He can
be contac
ted at ema
il: 2022201126@isiswa.uitm.edu.my.
Sofianita
Mutalib
is
currently
the
associate
professor
in
School
of
Computing
Scienc
es,
College
of
Computing
,
Infor
matics
and
Mathem
atics
Univ
ersiti
Teknol
ogi
MARA
,
(UiTM)
Shah
Alam
.
She
teaches
bachelor
and
postgraduate
courses
rela
ted
to
intelligent
systems
such
as
intelligent
system
development,
data
mining
,
and
final
project.
Her
primary
research
interests
involve
intelligent
systems,
data
mining
as
well
as
machin
e
learning
and
also
data
science. She c
an be contac
ted at email: sofia
nita@
uitm.edu.my
.
Syarifa
h Ruzaini
Syed Ar
is
is curren
tly a senior
lecture
r in School of
Computing
Scienc
es,
College
of
Computing
,
Infor
matics
and
Mathem
atics
Univ
ersiti
Teknol
ogi
MARA
(UiTM)
,
Shah Al
am. He
r
prima
ry re
searc
h
inter
ests inv
olve st
rateg
ic
manage
ment inf
ormatio
n
systems and business intellige
nce. She ca
n be contacte
d at email: ruza
ini@
uitm.edu.my
.
Evaluation Warning : The document was created with Spire.PDF for Python.