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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
, pp.
4642
~
4653
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
13
.i
4
.pp
4642
-
4653
4642
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
E
n
h
a
n
c
i
n
g t
e
xt
c
l
ass
i
f
i
c
at
i
on
t
h
r
ou
gh
n
ove
l
d
e
e
p
l
e
ar
n
i
n
g
se
q
u
e
n
t
i
al
at
t
e
n
t
i
on
f
u
si
on
ar
c
h
i
t
e
c
t
u
r
e
S
h
il
p
a, S
h
r
i
d
e
vi
S
om
a
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd E
ngi
ne
e
r
i
ng,
P
D
A
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
n
g
,
V
i
s
ve
s
va
r
a
ya
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
i
t
y,
K
a
l
a
bur
a
gi
,
I
ndi
a
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
N
ov 24, 2023
R
e
vi
s
e
d
M
a
r
8, 2024
A
c
c
e
pt
e
d
M
a
r
21, 2024
Text
classification
is
a
pivotal
task
within
natural
language
proc
essing
(NLP),
aimed
at
assigning
semantic
labels
to
text
sequences.
Trad
itional
methods
of
text
represe
ntation
often
fall
short
in
capturing
intrica
cies
in
contextu
al
informat
ion,
relying
heavily
on
manual
feature
extracti
on.
To
overcome
these
limitations,
this
research
work
presents
the
seq
uential
attenti
on
fusion
architectu
re
(
SAFA
)
to
enhance
the
features
extr
action.
SAFA
combines
deep
long
sort
-
term
memory
(
LSTM
)
and
mult
i
-
head
attenti
on
mechanism
(MHAM).
This
model
efficie
ntly
preserv
es
data
,
even
for
longer
phrases,
while
enhancing
local
attribute
understa
nding.
Additionally
, we introduce a
unique attenti
on mechanism
that optim
iz
es data
preservation,
a
crucial
element
in
text
classific
ation.
The
paper
also
o
utlines
a
comprehensive
framework,
incorporating
convolutional
layers
and
p
ooling
techniques,
designed
to
improve
feature
represe
ntation
and
e
nhance
classifi
cation
accuracy.
The
model'
s
effectivenes
s
is
demonst
rated
t
hrough
2
-
dimension
al
convolution
processes
and
advanced
pooling,
signif
icantly
improving
prediction
accura
cy.
This
resear
ch
not
only
contributes
to
the
development
of
more
accurate
text
classification
models
but
also
underscores the growing importance
of NLP techniques.
K
e
y
w
o
r
d
s
:
D
e
e
p l
e
a
r
ni
ng
M
ul
ti
-
h
e
a
d
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
N
a
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
T
e
xt
c
la
s
s
if
ic
a
ti
on
T
e
xt
r
e
pr
e
s
e
nt
a
ti
on
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
hi
lp
a
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g, P
D
A
C
ol
le
ge
of
E
ngi
ne
e
r
in
g
V
is
ve
s
va
r
a
y
a
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
it
y
K
a
la
bur
a
gi
, I
ndi
a
E
m
a
il
:
s
hi
lp
a
_122023@
r
e
di
f
f
m
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
A
c
c
or
di
ng
to
S
ta
ti
s
ti
c
s
in
2022
f
or
e
c
a
s
ts
,
it
is
pr
oj
e
c
te
d
th
a
t
th
e
gl
oba
l
num
be
r
of
in
te
r
ne
t
us
e
r
s
w
il
l
r
e
a
c
h
a
ppr
oxi
m
a
te
ly
5.3
bi
ll
io
n
by
2023
[
1]
.
T
he
di
gi
ta
l
s
phe
r
e
f
a
c
il
it
a
te
s
th
e
tr
a
n
s
f
e
r
of
va
r
io
us
f
or
m
s
of
c
ont
e
nt
,
pr
im
a
r
il
y
te
xt
ua
l
m
a
te
r
ia
l,
th
r
ough
th
e
p
r
oc
e
s
s
e
s
o
f
dow
nl
oa
di
ng
a
nd
upl
oa
di
ng.
E
a
c
h
da
y
a
s
s
ubs
ta
nt
ia
l
vol
um
e
of
in
f
or
m
a
ti
on
is
b
e
in
g
ge
ne
r
a
t
e
d,
it
i
s
vi
ta
l
to
e
nga
ge
in
m
a
nua
l
a
ll
ot
m
e
nt
of
a
s
ig
ni
f
ic
a
nt
qua
nt
it
y
of
te
xt
ua
l
c
ont
e
nt
,
how
e
ve
r
th
is
pr
oj
e
c
t
m
a
y
in
c
ur
s
ig
ni
f
ic
a
nt
c
os
ts
,
r
e
qui
r
e
a
s
ub
s
ta
nt
ia
l
ti
m
e
c
om
m
it
m
e
nt
,
a
nd
pr
e
s
e
nt
va
r
io
us
c
h
a
ll
e
nge
s
if
unde
r
ta
ke
n
[
2]
.
T
he
r
e
c
om
m
e
nd
e
d
a
ppr
oa
c
h
e
s
in
vol
ve
e
na
bl
in
g
te
xt
m
in
in
g
ope
r
a
ti
ons
a
nd
a
ut
om
a
ti
ng
in
f
o
r
m
a
ti
on
c
l
a
s
s
if
ic
a
ti
on.
T
e
xt
u
a
l
da
ta
is
a
va
lu
a
bl
e
s
our
c
e
of
in
f
or
m
a
ti
on,
hi
ghl
ig
ht
in
g
th
e
ne
c
e
s
s
it
y
of
e
m
pl
oyi
ng
e
f
f
ic
ie
nt
a
nd
c
os
t
-
e
f
f
e
c
ti
ve
te
c
hni
que
s
f
or
a
ut
om
a
te
d
or
ga
ni
z
a
ti
on
a
nd
a
n
a
ly
s
is
of
te
xt
s
in
a
c
a
d
e
m
ic
a
nd
c
om
m
e
r
c
ia
l
e
nvi
r
onm
e
nt
s
.
T
h
e
pr
im
a
r
y
obj
e
c
ti
ve
of
te
xt
c
la
s
s
if
ic
a
ti
on
is
to
a
s
s
ig
n
a
gi
ve
n
te
xt
to
a
pr
e
de
f
in
e
d
c
a
te
gor
y.
T
he
di
s
c
ip
li
ne
of
na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
(
N
L
P
)
e
nc
om
pa
s
s
e
s
a
w
id
e
r
a
nge
of
a
ppl
ic
a
ti
ons
,
th
e
r
e
by
in
vol
vi
ng
num
e
r
ous
r
e
s
pons
ib
il
it
ie
s
.
T
e
xt
c
a
te
gor
iz
a
ti
on
is
c
om
m
onl
y
di
vi
de
d
in
to
th
r
e
e
m
a
in
c
a
te
gor
ie
s
:
s
e
m
i
-
s
upe
r
vi
s
e
d,
un
s
upe
r
vi
s
e
d,
a
nd
s
upe
r
vi
s
e
d
[
3]
.
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
E
nhanc
in
g t
e
x
t
c
la
s
s
if
ic
at
io
n t
hr
ough nov
e
l
d
e
e
p l
e
ar
ni
ng
s
e
qu
e
nt
ia
l
at
te
nt
io
n f
us
io
n ar
c
hi
te
c
tu
r
e
(
Shi
lp
a)
4643
T
e
xt
c
l
a
s
s
if
ic
a
ti
on
is
an
e
s
s
e
nt
ia
l
s
ubt
a
s
k
w
it
hi
n
th
e
f
ie
ld
of
N
L
P
.
I
ts
m
a
in
goa
l
i
s
to
a
s
s
ig
n
one
or
m
or
e
la
be
ls
th
a
t
r
e
pr
e
s
e
nt
th
e
s
e
m
a
nt
ic
m
e
a
ni
ng
of
a
gi
ve
n
s
e
que
nc
e
of
te
xt
.
N
a
tu
r
a
l
la
ngua
ge
in
f
e
r
e
nc
e
(
N
L
I
)
e
nc
om
pa
s
s
e
s
th
r
e
e
ty
pe
s
of
c
la
s
s
if
ic
a
ti
on
ta
s
k
s
:
bi
na
r
y
c
la
s
s
if
ic
a
ti
on,
m
ul
ti
-
c
la
s
s
if
ic
a
ti
on,
a
nd
m
ul
ti
-
la
be
l
c
la
s
s
if
ic
a
ti
on.
T
he
a
f
or
e
m
e
nt
io
ne
d
e
nt
it
y
c
ons
ti
t
ut
e
s
a
n
in
te
gr
a
l
p
a
r
t
of
th
e
c
om
pr
e
he
n
s
iv
e
f
r
a
m
e
w
or
k.
F
ur
th
e
r
m
or
e
,
th
e
s
ys
te
m
in
c
or
por
a
te
s
va
r
io
us
m
odul
e
s
to
f
a
c
il
it
a
te
di
f
f
e
r
e
nt
f
unc
ti
ona
li
ti
e
s
.
T
he
s
e
m
odul
e
s
in
c
lu
de
que
s
ti
on
a
ns
w
e
r
in
g
(
Q
A
)
,
to
pi
c
c
l
a
s
s
if
ic
a
ti
on
(
T
C
)
,
ne
w
s
c
la
s
s
if
ic
a
ti
on
(
C
A
)
,
s
e
nt
im
e
nt
a
na
ly
s
i
s
(
S
A
)
, a
nd
ne
w
s
c
la
s
s
if
ic
a
ti
on
[
4]
.
T
he
ut
il
iz
a
ti
on
of
te
xt
r
e
pr
e
s
e
nt
a
ti
on
is
a
n
e
s
s
e
nt
ia
l
in
te
r
m
e
di
a
r
y
s
te
p
in
th
e
te
xt
c
la
s
s
if
ic
a
ti
on
pr
oc
e
dur
e
th
a
t
is
vi
ta
l
to
e
m
pl
oy.
T
he
s
ig
ni
f
ic
a
nc
e
of
w
or
d
c
o
nt
e
xt
w
it
hi
n
s
e
nt
e
nc
e
s
is
of
te
n
di
s
r
e
ga
r
de
d
by
c
onve
nt
io
na
l
m
e
th
ods
of
te
xt
r
e
pr
e
s
e
nt
a
ti
on
[
5]
,
f
ur
th
e
r
m
or
e
,
th
e
a
f
or
e
m
e
nt
io
ne
d
te
c
hni
que
s
a
r
e
c
ha
r
a
c
te
r
iz
e
d
by
th
e
ir
la
bour
-
in
te
ns
iv
e
na
tu
r
e
a
nd
r
e
qui
r
e
a
s
ubs
ta
nt
ia
l
a
ll
oc
a
ti
on
of
r
e
s
our
c
e
s
.
T
he
m
a
in
c
ont
r
ib
ut
in
g
f
a
c
to
r
to
th
is
is
s
ue
is
th
e
ir
r
e
li
a
nc
e
on
hu
m
a
n
op
e
r
a
to
r
s
to
pe
r
f
o
r
m
m
a
nua
l
f
e
a
tu
r
e
e
xt
r
a
c
ti
on.
T
he
us
e
r
'
s
w
r
it
te
n
c
ont
e
nt
e
xhi
bi
ts
l
a
c
k
of
c
la
r
it
y
a
nd
in
f
or
m
a
ti
on;
te
xt
c
la
s
s
if
ic
a
ti
on
m
ode
ls
a
r
e
c
on
s
tr
uc
te
d
us
in
g
th
e
pr
in
c
ip
le
s
of
te
xt
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng.
T
he
te
xt
ua
l
r
e
pr
e
s
e
nt
a
ti
ons
th
a
t
ha
ve
be
e
n
obt
a
in
e
d
c
a
n
be
e
f
f
e
c
ti
ve
ly
ut
il
iz
e
d
f
or
pr
e
c
is
e
te
xt
c
la
s
s
if
ic
a
ti
on,
gi
ve
n
th
a
t
th
e
y
pos
s
e
s
s
th
e
a
bi
li
ty
to
di
f
f
e
r
e
nt
ia
te
ba
s
e
d
on c
la
s
s
e
s
[
6]
.
C
onve
nt
io
na
l
a
ppr
oa
c
h
e
s
f
or
r
e
pr
e
s
e
nt
in
g
te
xt
,
s
uc
h
a
s
th
e
ve
c
to
r
s
pa
c
e
m
ode
l
[
7]
,
r
e
qui
r
e
th
e
ut
il
iz
a
ti
on
of
de
li
be
r
a
te
ly
c
r
a
f
te
d
f
e
a
tu
r
e
s
.
C
onv
e
nt
io
na
l
te
xt
r
e
pr
e
s
e
nt
a
ti
on
a
ppr
oa
c
he
s
ha
ve
ga
r
ne
r
e
d
s
ig
ni
f
ic
a
nt
a
tt
e
nt
io
n
due
to
th
e
ir
us
e
r
-
f
r
ie
ndl
y
na
tu
r
e
.
T
he
dr
a
w
ba
c
ks
of
c
e
r
ta
in
te
xt
r
e
pr
e
s
e
nt
a
ti
ons
in
c
lu
de
th
e
ir
hi
gh
di
m
e
ns
io
na
li
ty
a
nd
s
pa
r
s
it
y.
I
n
r
e
c
e
nt
ye
a
r
s
,
th
e
r
e
ha
s
be
e
n
not
a
bl
e
a
dv
a
nc
e
m
e
nt
in
th
e
dom
a
in
of
N
L
P
th
r
ough
th
e
a
ppl
ic
a
ti
on
of
de
e
p
l
e
a
r
ni
ng
te
c
hni
que
s
.
A
s
a
r
e
s
ul
t,
num
e
r
ous
m
ode
ls
ha
ve
be
e
n
d
e
ve
lo
pe
d
to
e
xt
r
a
c
t
pr
ope
r
ti
e
s
f
r
om
te
xt
.
T
he
m
ode
ls
ut
il
iz
e
d
in
th
is
s
tu
d
y
a
r
e
th
e
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
a
nd
th
e
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
k
(
R
N
N
)
[
8]
.
T
he
s
e
m
ode
ls
a
r
e
de
r
iv
e
d
f
r
om
c
la
s
s
ic
a
l
n
e
ur
a
l
ne
twor
ks
. T
he
s
e
two
m
e
th
ods
a
ddr
e
s
s
th
e
pr
obl
e
m
of
s
pa
r
s
e
,
hi
gh
-
di
m
e
ns
io
n
a
l
te
xt
r
e
pr
e
s
e
nt
a
ti
ons
by
e
m
pl
oyi
ng
a
n
e
nd
-
to
-
e
nd
le
a
r
ni
ng
pr
oc
e
s
s
f
or
te
xt
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
ons
.
H
ow
e
ve
r
,
th
e
ir
c
a
pa
bi
li
ty
to
c
ol
le
c
t
gl
oba
l
w
or
d
c
o
-
oc
c
ur
r
e
nc
e
s
i
n c
or
por
a
w
it
h non
-
c
ont
in
uous
a
nd l
ong
-
r
a
nge
s
e
m
a
nt
ic
s
i
s
l
im
it
e
d.
A
tt
e
nt
io
n
pr
oc
e
s
s
e
s
h
a
ve
be
e
n
w
id
e
ly
e
m
pl
oye
d
in
num
e
r
ous
s
tu
di
e
s
to
e
nha
nc
e
th
e
pe
r
f
or
m
a
nc
e
of
te
xt
c
la
s
s
if
ic
a
ti
on
m
ode
l
s
.
T
h
e
obj
e
c
ti
ve
of
th
e
s
e
a
ppr
oa
c
he
s
is
to
im
pr
ove
te
xt
r
e
pr
e
s
e
nt
a
ti
ons
by
in
c
or
por
a
ti
ng
c
om
pr
e
he
ns
iv
e
te
xt
s
e
m
a
nt
ic
s
.
T
he
m
a
jo
r
it
y
of
c
ur
r
e
nt
ne
ur
a
l
-
ba
s
e
d
te
xt
c
la
s
s
if
ic
a
ti
on
m
ode
ls
f
a
il
to
c
ons
id
e
r
th
e
in
te
r
a
c
ti
ons
be
twe
e
n
phr
a
s
e
s
in
a
te
xt
w
he
n
ge
ne
r
a
ti
ng
th
e
ir
te
xt
r
e
pr
e
s
e
nt
a
ti
ons
.
I
n
or
de
r
to
ove
r
c
om
e
th
e
a
f
or
e
m
e
nt
io
ne
d
li
m
it
a
ti
on,
te
xt
c
la
s
s
if
ic
a
ti
o
n
a
lg
or
it
hm
s
ha
ve
r
e
s
or
te
d
to
ut
il
iz
in
g
gr
a
ph
ne
ur
a
l
ne
twor
ks
(
G
N
N
s
)
a
s
a
pot
e
nt
ia
l
r
e
m
e
dy.
G
r
a
ph ne
ur
a
l
n
e
twor
ks
, c
om
m
onl
y r
e
f
e
r
r
e
d t
o
a
s
G
N
N
s
, ha
ve
de
m
ons
tr
a
te
d
s
ig
ni
f
ic
a
nt
a
dva
nc
e
m
e
nt
s
a
c
r
os
s
va
r
io
us
dom
a
i
ns
.
I
n
or
de
r
to
e
f
f
e
c
ti
ve
ly
ga
th
e
r
th
e
ove
r
a
ll
in
f
or
m
a
ti
on of
node
s
, ne
twor
ks
e
m
pl
oy a
m
e
s
s
a
ge
t
r
a
ns
m
is
s
io
n m
e
c
ha
ni
s
m
on gr
a
phs
. T
he
pr
im
a
r
y di
f
f
ic
ul
ty
a
s
s
oc
ia
t
e
d
w
it
h
th
e
s
e
m
e
th
od
s
pe
r
ta
in
s
to
th
e
de
v
e
lo
pm
e
nt
of
s
ui
ta
bl
e
te
xt
ua
l
gr
a
ph
s
.
T
he
T
e
xt
G
C
N
m
ode
l,
a
s
de
s
c
r
ib
e
d i
n
[
9]
, i
s
a
G
N
N
a
r
c
hi
te
c
tu
r
e
de
s
ig
ne
d s
pe
c
if
ic
a
ll
y f
or
t
e
xt
c
a
te
gor
iz
a
ti
on t
a
s
ks
.
T
h
e
o
pe
r
a
t
io
n
a
l
pr
o
c
e
dur
e
of
t
hi
s
a
ppr
oa
c
h
i
nv
ol
v
e
s
th
e
c
on
ve
r
s
i
on
of
t
e
xt
u
a
l
i
nf
or
m
a
t
io
n
in
t
o
a
c
om
pr
e
h
e
n
s
iv
e
h
e
t
e
r
o
ge
ne
ou
s
gr
a
p
h.
T
h
e
gr
a
ph
i
n
th
is
i
ll
u
s
tr
a
ti
on
de
pi
c
t
s
w
or
d
s
a
n
d
do
c
u
m
e
nt
s
a
s
di
s
ti
n
c
t
nod
e
s
,
pr
o
vi
d
in
g
a
r
e
pr
e
s
e
n
ta
t
io
n t
h
a
t
i
s
d
iv
i
de
d i
nt
o
t
w
o l
e
ve
l
s
o
f
de
ta
il
.
T
h
e
s
ub
s
e
q
ue
nt
ph
a
s
e
of
t
h
e
pr
oc
e
d
ur
e
in
vo
lv
e
s
th
e
ut
il
iz
a
t
io
n
of
a
gr
a
p
h
n
e
ur
a
l
n
e
tw
or
k
f
or
t
he
pur
p
os
e
of
c
la
s
s
if
yi
n
g
d
oc
um
e
nt
no
de
s
.
I
n
or
d
e
r
t
o
e
n
s
ur
e
c
on
s
is
te
nt
r
e
pr
e
s
e
nt
a
ti
on
l
e
a
r
ni
ng
f
or
do
c
u
m
e
nt
s
w
it
h
in
t
he
s
a
m
e
c
l
a
s
s
a
nd
d
if
f
e
r
e
nt
i
a
t
e
d
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng
f
or
d
oc
um
e
nt
s
a
c
r
os
s
di
f
f
e
r
e
n
t
c
l
a
s
s
e
s
,
i
t
i
s
e
s
s
e
nt
ia
l
t
o
o
bt
a
in
w
or
d
r
e
pr
e
s
e
nt
a
ti
on
s
a
n
d
e
f
f
e
c
ti
ve
l
y
di
s
tr
i
but
e
w
or
d
i
nf
or
m
a
t
io
n
a
m
o
ng
t
he
pa
pe
r
s
.
H
o
w
e
ve
r
,
it
i
s
i
m
por
t
a
nt
to
n
ot
e
th
a
t
a
s
i
ngl
e
w
or
d
c
a
n
po
s
s
e
s
s
m
ul
t
ip
l
e
m
e
a
ni
ng
s
, w
hi
c
h
c
a
n va
r
y
d
e
p
e
n
di
n
g
o
n
th
e
s
p
e
c
if
i
c
c
o
nt
e
xt
i
n
w
hi
c
h
i
t
i
s
e
m
p
lo
y
e
d
.
I
n t
h
e
c
ont
e
xt
of
te
c
hn
ol
o
gy
a
nd
f
ood
,
th
e
te
r
m
"
a
ppl
e
"
c
a
n
b
e
in
te
r
pr
e
te
d
in
t
w
o
di
s
t
in
c
t
w
a
y
s
.
T
he
t
e
r
m
"
i
t"
r
e
f
e
r
s
to
b
ot
h
t
he
f
r
ui
t
c
om
m
on
ly
k
no
w
n
a
s
a
pp
le
a
n
d
th
e
r
e
no
w
n
e
d
m
ul
ti
na
ti
on
a
l
t
e
c
hn
ol
ogy
c
o
m
p
a
ny
,
A
p
pl
e
I
n
c
.
T
h
e
a
f
or
e
m
e
nt
io
ne
d
p
hr
a
s
e
s
w
il
l
be
a
s
s
i
gn
e
d
t
o
d
is
ti
n
c
t
do
c
u
m
e
nt
c
a
te
gor
ie
s
,
r
e
s
ul
ti
ng
in
th
e
di
s
s
e
m
i
na
ti
o
n
of
va
r
i
ou
s
t
yp
e
s
of
in
f
or
m
a
ti
on a
c
r
o
s
s
t
he
s
e
d
oc
um
e
nt
s
. T
h
e
a
f
or
e
m
e
n
ti
o
ne
d
f
a
c
to
r
w
il
l
s
ig
ni
f
i
c
a
nt
ly
i
nf
l
ue
nc
e
t
h
e
m
a
n
ne
r
i
n w
hi
c
h
t
a
s
k
s
s
u
c
h
a
s
t
e
xt
c
la
s
s
if
i
c
a
ti
on
a
r
e
e
x
e
c
ut
e
d
i
n
s
u
bs
e
qu
e
nt
s
ta
ge
s
[
1
0]
.
T
he
e
s
c
a
la
ti
ng
gr
ow
th
of
di
gi
ta
l
c
ont
e
nt
,
pa
r
ti
c
ul
a
r
ly
in
te
xt
ua
l
f
or
m
on
th
e
in
te
r
ne
t,
unde
r
s
c
or
e
s
th
e
ne
e
d
f
or
r
obus
t
a
nd
e
f
f
ic
ie
nt
m
e
th
ods
of
te
xt
c
la
s
s
if
ic
a
ti
on
to
ha
ndl
e
th
is
va
s
t
a
nd
un
s
tr
uc
tu
r
e
d
da
ta
.
A
ut
om
a
ti
on
of
te
xt
c
la
s
s
if
ic
a
ti
on
ha
s
be
c
om
e
a
n
e
s
s
e
nt
ia
l
r
e
qui
r
e
m
e
nt
due
to
th
e
im
pr
a
c
ti
c
a
li
ty
a
nd
hi
gh c
os
t
of
m
a
nua
l
c
ont
e
nt
A
ll
ot
m
e
nt
.
T
h
e
pr
im
a
r
y
m
ot
iv
a
ti
on
be
hi
nd
th
is
r
e
s
e
a
r
c
h
s
te
m
s
f
r
om
th
e
de
m
a
nd
f
or
a
dva
nc
e
d
te
c
hni
que
s
in
th
e
f
ie
ld
of
N
L
P
to
de
ve
lo
p
m
or
e
pr
e
c
is
e
a
nd
e
f
f
ic
ie
nt
m
e
c
ha
ni
s
m
s
f
or
te
xt
c
la
s
s
if
ic
a
ti
on.
T
hi
s
s
tu
dy
is
dr
iv
e
n
by
th
e
r
e
c
ogni
ti
on
of
N
L
P
'
s
pot
e
nt
ia
l,
c
oupl
e
d
w
it
h
e
xi
s
ti
ng
m
e
th
odol
ogi
e
s
s
uc
h
a
s
gr
a
ph
ne
ur
a
l
ne
twor
ks
a
nd
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
,
w
hi
c
h
s
ig
ni
f
ic
a
nt
ly
e
nha
n
c
e
th
e
r
e
pr
e
s
e
nt
a
ti
on
of
te
xt
ua
l
da
ta
.
I
t
is
e
nvi
s
io
ne
d
th
a
t
th
e
s
e
a
d
va
nc
e
m
e
nt
s
w
il
l
not
onl
y
pr
ove
va
lu
a
bl
e
f
or
a
c
a
de
m
ic
r
e
s
e
a
r
c
h
but
a
ls
o
yi
e
ld
s
ubs
ta
nt
ia
l
a
dva
nt
a
ge
s
f
or
bus
in
e
s
s
e
s
na
vi
ga
ti
ng
th
e
e
r
a
of
in
f
or
m
a
ti
on
a
bunda
nc
e
.
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4642
-
4653
4644
‒
I
nnova
ti
ve
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
m
e
c
h
a
ni
s
m
(
M
H
A
M
)
-
ba
s
e
d
a
ppr
oa
c
h
:
th
is
r
e
s
e
a
r
c
h
pr
e
s
e
nt
s
a
n
in
nova
ti
ve
a
ppr
oa
c
h
to
te
xt
c
la
s
s
if
ic
a
ti
on
by
c
om
bi
ni
ng
M
H
A
M
m
ode
ls
w
it
h
d
eep
lo
ng
s
hor
t
-
te
r
m
m
e
m
or
y
(
L
S
T
M
)
.
T
he
pr
opos
e
d
m
ode
l
a
ddr
e
s
s
e
s
th
e
li
m
it
a
ti
ons
of
tr
a
di
ti
ona
l
M
H
A
M
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
by
e
f
f
e
c
ti
ve
ly
c
a
pt
ur
in
g l
oc
a
l
a
tt
r
ib
ut
e
s
a
nd e
nha
n
c
in
g t
he
unde
r
s
ta
ndi
ng of
t
e
xt
ua
l
da
ta
.
‒
E
f
f
ic
ie
nt
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
:
th
e
s
tu
dy
in
tr
oduc
e
s
a
nove
l
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
th
a
t
ov
e
r
c
om
e
s
th
e
li
m
it
a
ti
ons
of
s
e
lf
-
a
tt
e
nt
io
n
in
M
H
A
M
m
ode
ls
.
B
y
us
in
g
a
s
in
g
le
-
s
e
gm
e
nt
a
tt
e
nt
io
n
m
ode
l
in
c
onj
unc
ti
on
w
it
h
d
e
e
p,
it
e
f
f
ic
ie
nt
ly
c
a
pt
ur
e
s
d
e
pe
nde
nc
i
e
s
b
e
twe
e
n
te
r
m
s
a
nd
opt
im
iz
e
s
d
a
ta
pr
e
s
e
r
va
ti
on,
e
ve
n
f
or
lo
nge
r
phr
a
s
e
s
.
‒
A
dva
nc
e
d
te
xt
c
la
s
s
if
ic
a
ti
on
f
r
a
m
e
w
or
k
:
th
e
r
e
s
e
a
r
c
h
of
f
e
r
s
a
c
om
pr
e
he
ns
iv
e
te
xt
c
la
s
s
if
ic
a
ti
on
f
r
a
m
e
w
or
k t
ha
t
le
ve
r
a
ge
s
a
dva
nc
e
d t
e
c
hni
que
s
, i
nc
lu
di
ng a
tt
e
nt
io
n m
e
c
ha
ni
s
m
s
, c
onvolut
io
na
l
la
ye
r
s
, a
nd
pool
in
g,
to
im
pr
ove
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
T
hi
s
m
e
th
odol
ogy
c
ont
r
ib
ut
e
s
to
th
e
de
ve
lo
pm
e
nt
of
m
or
e
e
f
f
e
c
ti
ve
m
ode
ls
f
or
t
e
xt
c
la
s
s
if
ic
a
ti
o
n t
a
s
ks
.
‒
E
nha
nc
e
d
pr
e
di
c
ti
on
a
c
c
ur
a
c
y
:
th
e
pr
opos
e
d
m
e
th
odol
ogy
in
c
or
por
a
te
s
2
-
di
m
e
ns
io
na
l
c
onvolut
io
n
pr
oc
e
s
s
e
s
a
nd
a
dva
n
c
e
d
pool
in
g
te
c
hni
que
s
to
e
nha
nc
e
th
e
pr
e
di
c
ti
on
a
c
c
ur
a
c
y
of
te
xt
c
la
s
s
if
ic
a
ti
on.
B
y
e
f
f
ic
ie
nt
ly
e
xt
r
a
c
ti
ng
f
e
a
tu
r
e
s
a
nd
r
e
duc
in
g
th
e
c
om
pl
e
xi
ty
of
c
om
put
a
ti
ons
,
th
e
m
od
e
l
s
ig
ni
f
ic
a
nt
ly
im
pr
ove
s
i
ts
a
bi
li
ty
t
o a
c
c
ur
a
te
ly
c
la
s
s
if
y t
e
xt
da
ta
, m
a
ki
ng i
t
a
va
lu
a
bl
e
c
ont
r
ib
ut
io
n t
o t
he
f
ie
ld
of
N
L
P
.
T
he
r
e
s
e
a
r
c
h
or
ga
ni
z
a
ti
on
of
th
is
pa
pe
r
is
in
th
e
f
ir
s
t
s
e
c
ti
on
a
br
ie
f
ove
r
vi
e
w
is
gi
ve
n
of
th
e
te
xt
r
e
c
ogni
ti
on
.
I
n
th
e
s
e
c
ond
s
e
c
ti
on
a
th
or
ough
li
te
r
a
tu
r
e
s
ur
ve
y
i
s
gi
ve
n
of
th
e
e
xi
s
ti
ng
te
c
hni
qu
e
s
.
I
n
th
e
th
ir
d
s
e
c
ti
on
th
e
pr
opos
e
d
m
e
th
odol
ogy
is
gi
ve
n
w
he
r
e
in
th
e
nove
l
a
ppr
oa
c
h
is
de
s
ig
ne
d,
c
om
bi
ni
ng
a
M
H
A
M
a
tt
e
nt
io
n
m
ode
l
w
it
h
bi
di
r
e
c
ti
ona
l
L
S
T
M
,
a
nd
it
s
a
r
c
hi
te
c
tu
r
e
f
or
te
xt
c
la
s
s
if
ic
a
ti
on.
I
n
th
e
la
s
t
s
e
c
ti
on
th
e
pe
r
f
or
m
a
nc
e
a
na
ly
s
is
i
s
gi
ve
n
w
hi
c
h di
s
pl
a
y
s
t
he
r
e
s
ul
ts
i
n f
or
m
of
t
a
bl
e
s
a
nd gr
a
phs
.
2.
R
E
L
A
T
E
D
WORK
F
or
th
e
pur
pos
e
of
te
xt
c
a
te
gor
iz
a
ti
on
two
c
a
te
gor
ie
s
a
r
e
of
te
n
us
e
d
in
pr
e
vi
ous
a
ut
om
a
te
d
s
ys
te
m
s
.
P
a
tt
e
r
n
m
a
tc
hi
ng
c
a
n
be
a
c
c
om
pl
is
he
d
us
in
g
one
o
f
two
di
f
f
e
r
e
nt
m
e
th
ods
.
M
a
c
hi
ne
le
a
r
ni
ng
m
ode
ls
a
nd
r
ul
e
-
ba
s
e
d
m
a
nua
l
la
be
ll
in
g
a
r
e
th
e
two
m
a
in
a
ppr
oa
c
he
s
us
e
d
f
or
da
ta
la
be
ll
in
g.
R
ul
e
-
ba
s
e
d
ha
nd
-
c
r
a
f
t
la
be
ll
in
g
in
vol
ve
s
m
a
nua
ll
y
a
ppl
yi
ng
la
be
l
s
to
da
ta
a
c
c
or
di
ng
to
pr
e
de
f
in
e
d
r
ul
e
s
[
11]
.
M
a
c
hi
ne
l
e
a
r
ni
ng
m
ode
ls
,
on
th
e
ot
h
e
r
ha
nd,
us
e
pr
e
-
de
f
in
e
d
la
be
l
s
-
li
ke
n
a
ïv
e
B
a
ye
s
(
N
B
)
,
K
-
ne
a
r
e
s
t
ne
ig
hbour
s
(
K
N
N
)
,
a
nd
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
-
to
a
na
ly
s
e
pa
tt
e
r
ns
a
nd
c
or
r
e
la
ti
ons
in
th
e
da
ta
s
e
t
a
nd
a
s
s
ig
n
la
be
ls
to
da
ta
.
T
w
o
m
a
jo
r
is
s
ue
s
a
r
e
r
e
s
pon
s
ib
le
f
or
th
e
s
e
a
ppr
oa
c
he
s
in
a
b
il
it
y
to
a
c
hi
e
ve
hi
gh
a
c
c
ur
a
c
y
p
e
r
f
or
m
a
nc
e
.
B
e
c
a
us
e
th
e
da
ta
s
e
t
f
or
m
a
ts
w
e
r
e
f
ir
s
t
li
m
it
e
d
to
s
tr
uc
tu
r
e
d
da
t
a
,
unpr
oc
e
s
s
e
d
te
xt
da
ta
pr
e
s
e
nt
e
d
c
ha
ll
e
nge
s
f
or
a
lg
or
it
h
m
s
li
ke
K
N
N
a
nd
S
V
M
[
12
]
.
T
hi
s
m
e
a
nt
th
a
t
a
la
r
g
e
qua
nt
it
y
of
da
ta
t
r
a
ns
f
or
m
a
ti
on
w
a
s
ne
e
de
d,
w
hi
c
h
in
te
r
f
e
r
e
d
w
it
h
th
e
a
dm
in
is
tr
a
ti
on
o
f
in
dus
tr
ia
l
da
ta
a
nd
w
a
s
in
c
ons
i
s
te
nt
w
it
h
hum
a
n
c
ogni
ti
ve
pr
oc
e
s
s
e
s
.
F
ur
th
e
r
m
or
e
,
th
e
a
m
ount
of
th
e
da
ta
s
e
t
w
a
s
bi
gge
r
,
ne
c
e
s
s
it
a
ti
ng
th
e
pr
e
-
tr
a
in
in
g
of
te
c
hnol
ogi
e
s
li
ke
tr
a
ns
f
or
m
e
r
to
in
c
r
e
a
s
e
a
c
c
ur
a
c
y
in
da
ta
m
in
in
g
a
nd
m
a
c
hi
ne
le
a
r
ni
ng
ta
s
ks
.
T
h
e
r
e
f
or
e
,
th
e
e
f
f
e
c
ti
ve
ne
s
s
a
nd
pe
r
f
or
m
a
nc
e
of
e
xi
s
ti
ng
N
L
P
ta
s
ks
w
e
r
e
ha
m
pe
r
e
d
by
th
e
s
e
m
e
th
odol
ogi
e
s
.
C
om
pa
r
e
d
to
c
la
s
s
ic
a
l
m
ode
ls
th
a
t
ne
e
d
th
e
id
e
nt
if
ic
a
ti
on
of
f
unc
ti
ons
or
ta
s
k
s
f
o
r
f
or
e
c
a
s
ti
ng
or
c
la
s
s
if
ic
a
ti
on,
d
e
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
ha
ve
ga
in
e
d
im
por
ta
nc
e
in
th
e
f
ie
ld
o
f
N
L
P
f
o
r
te
xt
c
a
te
gor
iz
a
ti
on
or
ge
ne
r
a
ti
on
ta
s
ks
.
V
a
r
io
us
da
ta
s
e
ts
a
r
e
s
uppl
ie
d f
or
t
he
pr
oc
e
dur
e
s
[
13]
. T
he
ir
a
bi
li
ty
t
o c
om
bi
ne
s
e
ve
r
a
l
ta
s
k
s
i
nt
o a
s
in
gl
e
m
ode
l
-
w
hi
c
h
in
c
lu
de
s
pr
e
-
tr
a
in
e
d
la
ngu
a
ge
m
ode
l
s
li
ke
tr
a
ns
f
or
m
e
r
,
bi
di
r
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
(
B
E
R
T
)
, a
nd
ge
ne
r
a
ti
ve
pr
e
-
tr
a
in
e
d t
r
a
ns
f
or
m
e
r
s
(
G
P
T
)
-
is
t
he
s
our
c
e
of
t
hi
s
c
a
pa
bi
li
ty
[
14]
.
T
he
m
ul
ti
-
c
la
s
s
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
M
C
N
N
)
-
L
S
T
M
m
e
th
od
f
or
te
xt
c
la
s
s
if
ic
a
ti
on
in
ne
w
s
da
ta
i
s
pr
e
s
e
nt
e
d i
n t
hi
s
r
e
s
e
a
r
c
h. T
o a
c
hi
e
ve
t
he
de
s
ir
e
d r
e
s
ul
ts
, t
hi
s
m
e
th
od c
om
bi
ne
s
L
S
T
M
a
nd C
N
N
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
.
W
or
d
s
p
a
ti
a
l
or
ga
ni
z
a
ti
on
in
phr
a
s
e
s
,
pa
r
a
gr
a
phs
,
or
pa
ge
s
c
a
n
be
c
a
pt
ur
e
d
us
in
g
C
N
N
s
.
W
he
n
pr
oc
e
s
s
in
g
te
xt
-
ba
s
e
d
in
put
da
ta
,
th
e
y
a
r
e
c
om
m
onl
y
us
e
d
a
s
f
e
a
tu
r
e
e
xt
r
a
c
to
r
s
in
c
onj
unc
ti
on
w
it
h
ne
twor
ks
L
S
T
M
s
[
15]
.
T
hi
s
r
e
s
e
a
r
c
h
in
tr
oduc
e
s
a
ne
w
a
ppr
oa
c
h
c
a
ll
e
d
th
e
m
ul
ti
-
ha
s
hi
ng
e
m
be
ddi
ng
-
ba
s
e
d
di
f
f
e
r
e
nt
ia
l
ne
ur
a
l
a
r
c
hi
te
c
tu
r
e
s
e
a
r
c
h
te
c
h
ni
que
f
or
e
xpr
e
s
s
in
g
m
ul
ti
li
ngua
l
te
xt
.
T
h
e
pur
pos
e
of
bui
ld
in
g
a
m
ul
ti
-
ha
s
hi
ng
ne
twor
k
is
to
e
f
f
e
c
ti
ve
ly
tr
a
ns
f
e
r
s
ynt
a
c
ti
c
a
nd
c
ont
e
xt
ua
l
s
e
m
a
nt
ic
da
ta
be
twe
e
n
la
ngua
ge
s
[
16]
.
T
hi
s
ne
twor
k
is
a
l
s
o
bui
lt
to
ha
ndl
e
a
va
r
ie
ty
of
gr
a
ph
da
ta
f
or
m
a
ts
w
it
h
e
a
s
e
.
U
s
in
g
pa
r
a
m
e
te
r
iz
a
ti
on
-
ba
s
e
d
gr
a
di
e
nt
s
e
a
r
c
h,
ne
twor
k
to
pol
ogi
e
s
f
or
m
ul
ti
li
ngua
l
te
xt
r
e
pr
e
s
e
nt
a
ti
on
a
r
e
e
xpl
or
e
d.
F
or
e
a
c
h
c
a
ndi
da
te
ope
r
a
ti
on
in
th
e
s
e
a
r
c
h
s
pa
c
e
,
a
c
ont
in
uous
e
nc
ode
r
-
m
or
e
pr
e
c
is
e
ly
,
a
ne
ur
a
l
te
ns
or
ne
twor
k
w
it
h
m
ul
ti
-
ha
s
hi
ng
e
m
be
ddi
ng
-
is
e
m
pl
oye
d
to
e
s
ti
m
a
te
th
e
li
ke
li
hood.
M
a
xi
m
iz
in
g
s
e
a
r
c
h
pr
oc
e
s
s
e
f
f
ic
ie
nc
y i
s
t
he
a
im
of
t
hi
s
m
e
th
odol
ogy
[
17]
.
I
n
th
is
r
e
s
e
a
r
c
h,
a
uni
que
s
e
lf
-
s
upe
r
vi
s
e
d
a
tt
e
nt
io
n
m
e
th
od
is
pr
e
s
e
nt
e
d
th
a
t
doe
s
not
r
e
qui
r
e
a
ny
a
nnot
a
ti
on
c
os
t
s
by
us
in
g
p
e
r
tu
r
ba
ti
ons
to
s
uppor
t
a
tt
e
nt
io
n
le
a
r
ni
ng.
T
he
m
a
in
obj
e
c
ti
ve
is
to
im
pr
ove
a
c
c
ur
a
c
y
by
a
m
pl
if
yi
ng
th
e
noi
s
e
le
v
e
l
of
s
pe
c
if
ic
w
or
ds
in
th
e
gi
ve
n
s
e
nt
e
nc
e
.
P
r
e
s
e
r
vi
ng
th
e
phr
a
s
e
'
s
pr
e
di
c
ti
ve
pow
e
r
a
nd
m
e
a
ni
ng
dur
in
g
t
hi
s
pr
oc
e
dur
e
is
c
r
it
ic
a
l
[
18]
.
U
s
in
g
a
n
a
tt
e
nt
io
n
-
ba
s
e
d
ga
te
d
gr
a
ph
ne
ur
a
l
ne
twor
k
is
th
e
s
ugge
s
te
d
a
ppr
oa
c
h
f
or
th
e
a
ut
om
a
ti
c
e
xt
r
a
c
ti
on
of
node
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
ons
.
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
E
nhanc
in
g t
e
x
t
c
la
s
s
if
ic
at
io
n t
hr
ough nov
e
l
d
e
e
p l
e
ar
ni
ng
s
e
qu
e
nt
ia
l
at
te
nt
io
n f
us
io
n ar
c
hi
te
c
tu
r
e
(
Shi
lp
a)
4645
P
a
r
ti
c
ul
a
r
ly
de
s
ig
ne
d
f
or
th
e
c
ont
e
xt
of
c
onne
c
te
d
P
s
y
s
te
m
s
i
s
th
e
c
oupl
e
d
p
gr
a
ph
a
tt
e
nt
io
n
n
e
ur
a
l
ne
twor
k
(
C
P
G
A
N
N
)
te
c
hni
que
.
T
o
e
f
f
e
c
ti
ve
ly
r
e
duc
e
th
e
de
pe
nde
nc
y
o
n
non
-
s
e
que
nt
ia
l
w
or
ds
ove
r
e
xt
e
nde
d
pe
r
io
ds
of
ti
m
e
,
th
e
ga
ti
ng
uni
t
i
s
be
in
g
de
ve
lo
pe
d
w
it
h
a
tt
e
nt
io
n
to
c
a
pt
ur
e
s
e
m
a
nt
ic
li
nks
w
it
hi
n
th
e
c
ont
e
xt
.
T
he
a
tt
e
nt
io
n
te
c
hni
que
is
us
e
d
by
th
e
C
P
G
A
N
N
to
e
xt
r
a
c
t
ke
yw
o
r
d
node
s
.
T
he
node
s
a
bi
li
ty
to
di
s
ti
ngui
s
h
f
o
r
c
la
s
s
if
ic
a
ti
on
is
im
pr
ove
d
by
th
is
e
xt
r
a
c
ti
on
pr
oc
e
dur
e
,
s
u
bgr
a
ph
r
e
pr
e
s
e
nt
a
ti
ons
a
r
e
th
e
n
a
ggr
e
ga
te
d
th
r
oughout t
he
r
e
a
dout
pr
oc
e
s
s
us
in
g t
he
r
e
tr
ie
ve
d ke
yw
or
d nod
e
s
[
19]
.
O
pt
im
iz
in
g t
he
e
f
f
e
c
ti
ve
ne
s
s
a
nd a
c
c
ur
a
c
y of
r
e
la
ti
ve
di
s
c
r
im
in
a
ti
on c
r
it
e
r
io
n
(
R
D
C
)
f
e
a
tu
r
e
r
a
nki
n
g
is
t
he
m
a
in
obj
e
c
ti
ve
of
t
he
a
lt
e
r
na
ti
ve
r
e
la
ti
ve
di
s
c
r
im
in
a
ti
on c
r
it
e
r
io
n
(
A
R
D
C
)
. T
he
pr
im
a
r
y obje
c
ti
ve
of
t
h
e
A
R
D
C
is
to
s
e
e
k
a
nd
id
e
nt
if
y
w
or
ds
th
a
t
a
r
e
c
om
m
onl
y
us
e
d
i
n
th
e
pos
it
iv
e
c
la
s
s
.
T
he
R
D
C
a
nd
in
f
or
m
a
ti
on
r
e
la
ti
ve
di
s
c
r
im
in
a
ti
on
c
r
it
e
r
io
n
(
I
R
D
C
)
m
e
th
odol
ogi
e
s
'
out
c
o
m
e
s
w
e
r
e
c
om
pa
r
e
d,
a
nd
th
e
s
tu
dy
a
ls
o
lo
oke
d
a
t
be
nc
hm
a
r
ki
ng
f
unc
ti
ons
th
a
t
a
r
e
of
te
n
us
e
d,
s
uc
h
a
s
in
f
or
m
a
ti
on
ga
in
(
I
G
)
,
P
e
a
r
s
on
c
or
r
e
la
ti
on
c
oe
f
f
ic
ie
nt
(
P
C
C
)
,
a
nd
R
e
li
e
f
F
.
B
oo
s
t
e
xt
e
ns
ib
le
m
a
r
kup
la
ngua
ge
(
X
M
L
)
,
a
n
X
M
L
-
ba
s
e
d
m
e
th
od
f
or
e
xt
r
e
m
e
m
ul
ti
la
be
l
te
xt
c
a
te
gor
iz
a
ti
on,
is
pr
e
s
e
nt
e
d
in
th
is
w
or
k.
B
oos
tXM
L
is
a
de
e
p
le
a
r
ni
ng
f
r
a
m
e
w
or
k
th
a
t
ha
s
be
e
n
s
ig
ni
f
ic
a
nt
ly
im
pr
ove
d
vi
a
th
e
u
s
e
of
gr
a
di
e
nt
boos
ti
ng
t
e
c
hni
que
s
.
T
hr
ough
th
e
us
e
of
ta
il
l
a
be
ls
a
nd
a
pr
im
a
r
y
f
oc
us
on
unf
it
te
d
tr
a
in
in
g
in
s
ta
nc
e
s
,
th
e
B
oo
s
tXM
L
m
e
th
od
e
nha
nc
e
s
th
e
r
e
s
id
ua
l.
T
h
e
s
ugge
s
te
d
m
e
th
od
in
vol
ve
s
gi
vi
ng
ta
il
la
be
ls
m
or
e
w
e
ig
ht
s
a
t
e
ve
r
y
B
oo
s
ti
ng
S
te
p.
I
t'
s
be
s
t
to
in
c
lu
de
a
c
or
r
e
c
ti
ve
s
t
e
p
th
r
oughout
th
e
opt
im
iz
a
ti
on
pr
oc
e
s
s
to
ha
ndl
e
th
e
po
s
s
ib
le
is
s
ue
of
m
is
m
a
tc
hi
ng
b
e
twe
e
n
th
e
te
xt
e
nc
ode
r
a
nd w
e
a
k l
e
a
r
ne
r
s
.
T
he
m
ode
l
pe
r
f
or
m
s
be
tt
e
r
w
he
n t
he
r
e
i
s
l
e
s
s
c
ha
nc
e
of
r
unni
ng a
c
r
os
s
l
oc
a
l
opt
im
a
[
9]
.
3.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
O
L
O
G
Y
C
ons
id
e
r
in
g
th
e
tr
a
di
ti
ona
l
m
e
th
ods
of
M
H
A
M
a
tt
e
nt
io
n
m
ode
l
be
in
g
us
e
d,
th
e
m
e
c
ha
ni
s
m
of
s
e
lf
-
a
tt
e
nt
io
n
is
not
e
f
f
e
c
ti
ve
to
unde
r
s
ta
nd
a
nd
gr
a
s
p
th
e
lo
c
a
l
a
tt
r
ib
ut
e
s
f
r
om
th
e
ve
c
to
r
s
of
ph
r
a
s
e
th
a
t
e
ve
nt
ua
ll
y
om
it
s
e
s
s
e
nt
ia
l
d
a
ta
in
phr
a
s
e
s
.
A
ddi
ti
ona
ll
y,
th
e
c
om
pr
e
he
ns
iv
e
p
e
r
f
or
m
a
nc
e
by
th
e
M
H
A
M
a
tt
e
nt
io
n
m
ode
l
is
la
c
ki
ng
b
e
c
a
us
e
of
th
e
a
b
s
e
nc
e
of
la
ngua
g
e
m
ode
ll
in
g
c
a
pa
c
it
y.
H
e
nc
e
,
a
f
r
a
m
e
w
or
k
is
pr
opos
e
d f
or
t
he
M
H
A
M
a
tt
e
nt
io
n m
ode
l
a
nd t
he
a
r
c
hi
te
c
tu
r
e
o
f
t
hi
s
m
ode
l
is
gi
ve
n i
n t
he
F
ig
ur
e
1
.
F
ig
ur
e
1
.
T
he
a
r
c
hi
te
c
tu
r
e
of
t
he
pr
opos
e
d
M
H
A
M
a
tt
e
nt
io
n m
ode
l
T
hi
s
r
e
s
e
a
r
c
h
w
or
k
a
s
s
um
e
s
=
{
}
de
not
e
s
th
e
phr
a
s
e
s
e
t
f
or
e
v
e
r
y
phr
a
s
e
th
a
t
is
te
r
m
e
d
a
s
a
pa
r
ti
c
ul
a
r
c
la
s
s
,
th
e
in
put
da
ta
of
is
tr
a
ns
f
e
r
r
e
d
to
th
e
in
tr
oduc
e
d
m
e
th
od
to
de
r
iv
e
a
tt
r
ib
ut
e
s
.
O
nc
e
th
e
phr
a
s
e
ve
c
to
r
is
ge
ne
r
a
te
d
th
a
t
is
de
not
e
d
by
to
a
tt
a
in
w
hi
c
h
is
th
e
hi
dde
n
s
ta
te
.
F
u
r
th
e
r
,
th
e
a
tt
e
nt
io
n
m
e
th
od
is
us
e
d
to
gr
a
s
p
th
e
in
c
r
e
a
s
e
d
di
s
ta
nc
e
d
a
ta
f
r
om
us
in
g
th
e
s
e
te
r
m
s
th
a
t
ha
ve
in
c
r
e
a
s
e
d
w
e
ig
ht
s
.
H
ow
e
ve
r
,
a
pool
ne
twor
k
is
us
e
d
f
or
de
c
r
e
a
s
in
g
th
e
s
i
z
e
of
.
A
t
la
s
t,
th
e
f
in
a
l
e
xpr
e
s
s
io
n
is
pr
oduc
e
d
by
th
e
c
ont
r
a
c
te
d
ne
twor
k
th
a
t
ha
s
a
c
om
pl
e
te
ly
li
nke
d
la
ye
r
,
w
hi
c
h
i
s
ut
il
iz
e
d
f
or
pr
e
di
c
ti
on
of
th
e
phr
a
s
e
ta
g.
T
h
e
pr
oba
bi
li
ty
of
th
e
phr
a
s
e
f
it
s
in
th
e
pa
r
ti
c
ul
a
r
c
la
s
s
is
r
e
pr
e
s
e
nt
e
d
a
s
(
|
,
)
,
in
w
hi
c
h
th
e
ne
twor
k
a
tt
r
ib
ut
e
i
s
r
e
pr
e
s
e
nt
e
d a
s
.
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4642
-
4653
4646
3.1.
M
u
lt
i
-
h
e
ad
at
t
e
n
t
io
n
m
e
c
h
an
is
m
m
od
e
l
an
d
l
on
g s
h
or
t
-
t
e
r
m
m
e
m
o
r
y
T
he
m
ode
l
pr
opos
e
d
in
th
is
s
tu
dy
is
di
s
ti
nc
t
f
r
om
th
e
tr
a
di
ti
ona
l
m
e
th
ods
us
e
d
a
s
M
H
A
M
m
ode
l,
a
s
th
e
pr
opos
e
d
m
e
th
od
us
e
s
a
bi
di
r
e
c
ti
ona
l
L
S
T
M
to
e
xt
r
a
c
t
a
tt
r
i
but
e
s
th
a
t
om
it
s
th
e
de
pe
nde
n
c
y
is
s
u
e
c
a
u
s
e
d
by
R
N
N
a
s
w
e
ll
a
s
de
c
r
e
a
s
e
s
th
e
c
om
pl
e
xi
ty
of
c
om
put
a
ti
on
s
th
a
t
a
r
is
e
s
due
to
th
e
m
ul
ti
-
le
ve
l
a
tt
e
nt
io
n
m
ode
l.
A
lt
hough,
th
e
r
e
is
lo
s
s
of
da
ta
in
bi
di
r
e
c
ti
ona
l
L
S
T
M
w
hi
le
th
e
le
ngt
h
of
a
phr
a
s
e
pr
ol
ongs
a
pa
r
ti
c
ul
a
r
le
ngt
h.
S
ta
ti
ng
th
a
t
bi
di
r
e
c
ti
ona
l
L
S
T
M
doe
s
not
c
o
nt
r
ib
ut
e
to
m
odi
f
ic
a
ti
on
of
da
ta
pr
e
s
e
r
va
ti
on.
T
he
r
e
f
or
e
,
a
s
in
gl
e
s
e
gm
e
nt
a
tt
e
nt
io
n
m
ode
l
is
c
om
bi
ne
d
w
it
h
th
e
bi
di
r
e
c
ti
on
L
S
T
M
to
gr
a
s
p
va
r
io
us
f
e
a
tu
r
e
s
.
T
he
r
e
la
ti
on
be
twe
e
n
e
a
c
h
of
two
te
r
m
s
is
e
va
lu
a
te
d
us
in
g
ve
c
to
r
s
s
uc
h
a
s
ke
y
s
a
nd
que
r
ie
s
to
ga
in
th
e
i
nt
e
r
na
l
li
nks
of
t
he
phr
a
s
e
s
. H
e
n
c
e
, t
he
a
tt
e
nt
io
n m
ode
l
pr
o
pos
e
d e
f
f
ic
ie
nt
ly
om
it
s
a
ny l
os
s
of
da
ta
.
C
ons
id
e
r
a
gi
ve
n
phr
a
s
e
=
{
1
,
2
,
…
.
.
,
}
,
w
he
r
e
e
ve
r
y
te
r
m
is
tr
a
ns
f
or
m
e
d
a
s
a
ve
c
to
r
by
s
e
le
c
ti
ng a
m
a
tr
ix
th
a
t
i
s
e
m
b
e
dde
d
×
,
in
w
hi
c
h
th
e
w
or
d
s
i
z
e
th
a
t
is
e
m
b
e
dde
d
i
s
de
not
e
d
a
s
a
nd
th
e
di
m
e
ns
io
n
of
voc
a
bul
a
r
y
.
C
ons
id
e
r
=
∗
,
in
th
is
c
a
s
e
is
us
e
d
to
de
not
e
th
e
−
ℎ
te
r
m
in
th
e
ph
r
a
s
e
.
A
f
te
r
w
hi
c
h
is
tr
a
ns
f
e
r
r
e
d
to
th
e
e
nc
ode
d
p
os
it
io
n
ne
twor
k
to
a
tt
a
in
th
e
pos
it
io
n
da
ta
in
th
e
in
put
.
A
s
s
um
e
=
r
e
pr
e
s
e
nt
s
th
e
out
put
of
pos
t
th
e
e
nc
ode
d
pos
it
io
n.
T
he
e
qua
ti
on
s
f
or
e
nc
ode
d pos
it
io
n
a
r
e
f
or
m
ul
a
te
d a
s
gi
ve
n i
n (
1)
.
(
,
2
)
=
(
/
10000
2
/
)
(
1)
(
,
2
)
=
(
/
10000
2
/
)
(
2)
I
n
th
e
(
1)
a
nd
(
2
)
,
th
e
s
iz
e
of
ne
twor
k
is
de
not
e
d
a
s
a
nd
th
e
ve
c
to
r
s
iz
e
is
gi
ve
n
a
s
.
T
hi
s
r
e
s
e
a
r
c
h
w
or
k
us
e
s
th
e
bi
di
r
e
c
ti
ona
l
L
S
T
M
m
ode
l
th
a
t
ha
s
a
f
or
w
a
r
d
a
nd
ba
c
kw
a
r
d
m
e
th
od
to
pr
oduc
e
th
e
c
onc
e
a
le
d
r
e
pr
e
s
e
nt
a
ti
on.
T
he
f
or
w
a
r
d
L
S
T
M
r
e
c
e
iv
e
s
{
1
,
2
,
…
.
.
,
}
to
a
tt
a
in
th
e
f
or
w
a
r
d
c
on
c
e
a
le
d
e
xpr
e
s
s
io
n
{
0
,
1
,
…
.
.
,
}
by
e
nc
odi
ng,
th
e
ba
c
k
L
S
T
M
pr
oduc
e
s
a
ba
c
k
w
a
r
d
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
{
0
,
1
,
…
.
.
,
}
onc
e
th
e
ve
c
to
r
s
{
,
−
1
,
…
.
.
,
1
}
a
r
e
im
por
te
d.
L
a
s
tl
y,
th
e
f
or
w
a
r
d
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
a
s
w
e
ll
a
s
th
e
ba
c
kw
a
r
d
c
onc
e
a
l
e
d
e
xpr
e
s
s
io
n
s
a
r
e
s
pl
it
th
a
t
r
e
s
ul
ts
in
:
{
0
+
0
,
1
+
1
,
…
.
.
,
0
+
}
.
C
ons
id
e
r
in
g
th
e
be
ne
f
it
s
a
nd
c
ha
ll
e
nge
s
of
th
e
a
tt
e
nt
io
n
m
o
de
l,
a
n
in
di
vi
dua
l
he
a
d
a
tt
e
nt
io
n
m
ode
l
is
pr
opos
e
d
to
a
tt
a
in
e
s
s
e
nt
ia
l
da
ta
f
r
om
th
e
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
.
T
hi
s
r
e
s
e
a
r
c
h
w
or
k
ta
ke
in
to
a
c
c
ount
th
e
w
e
ig
ht
of
th
e
ot
he
r
te
r
m
s
dur
in
g
e
va
lu
a
ti
on
of
th
e
a
tt
e
nt
io
n
va
l
ue
of
a
s
in
gl
e
te
r
m
th
r
ough
a
nd
,
w
he
r
e
in
a
nd
a
r
e
th
e
r
e
s
ul
ti
ng
ve
c
to
r
s
of
th
e
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
,
th
e
l
in
k
be
twe
e
n
a
s
w
e
ll
a
s
is
th
e
a
tt
e
nt
io
n
va
lu
e
f
or
di
f
f
e
r
e
nt
te
r
m
s
to
th
e
te
r
m
,
(
1
,
)
.
W
hi
le
c
ons
id
e
r
in
g
th
e
a
tt
e
nt
io
n
m
e
th
od,
th
e
e
qua
ti
ons
a
r
e
gi
ve
n a
s
m
e
nt
io
ne
d i
n (
3)
a
nd (
4)
.
,
=
(
)
(
)
−
1
2
(
3)
,
′
=
(
,
)
(
∑
=
1
(
,
)
)
−
1
(
4
)
C
ons
id
e
r
in
g
th
e
(
3)
a
nd
(
4)
,
,
(
1
,
)
,
is
th
e
s
iz
e
of
a
nd
to
s
to
p
th
e
dot
pr
oduc
t
to
be
huge
.
T
he
(
4)
is
th
e
f
or
m
ul
a
ti
on
f
or
tr
a
ns
f
or
m
a
ti
on
of
a
tt
e
nt
io
n
va
lu
e
us
in
g
th
e
f
unc
ti
on
s
of
tm
a
x.
T
he
a
tt
e
nt
io
n
m
e
th
od
out
put
is
e
va
lu
a
te
d
us
in
g
th
e
e
qua
ti
on
gi
ve
n
i
n
(
5)
.
H
e
r
e
,
th
e
c
ount
of
te
r
m
s
is
de
not
e
d
a
s
f
or
th
e
phr
a
s
e
,
is
us
e
d
to
e
xpr
e
s
s
th
e
v
e
c
to
r
r
e
s
ul
te
d
f
r
om
.
T
he
s
um
of
w
e
ig
ht
s
f
r
om
va
r
io
us
in
put
s
i
s
ta
ke
n i
nt
o c
ons
id
e
r
a
ti
on by f
ix
in
g t
he
va
lu
e
of
,
′
.
′
=
∑
=
1
,
′
(
5)
3.2.
P
ool
n
e
t
w
or
k
T
hi
s
r
e
s
e
a
r
c
h
w
or
k
obs
e
r
ve
th
a
t
th
e
s
tr
uc
tu
r
e
of
th
e
a
tt
e
nt
io
n
m
ode
l
in
te
r
na
ll
y
ha
s
tr
a
ns
f
or
m
a
ti
ons
th
a
t
a
r
e
li
ne
a
r
,
th
is
una
voi
da
bl
y
unde
r
m
in
e
s
th
e
r
e
pr
e
s
e
nt
a
ti
o
n
of
e
ve
r
y
te
r
m
.
T
hi
s
is
is
s
ue
is
r
e
s
ol
ve
d
by
m
a
ppi
ng
th
e
e
xpr
e
s
s
io
n
f
r
om
s
m
a
ll
to
la
r
ge
s
iz
e
a
nd
f
ur
th
e
r
t
o
lo
w
s
iz
e
by
ut
il
iz
in
g
th
e
c
om
pl
e
te
ly
li
nke
d
la
ye
r
.
F
ur
th
e
r
m
o
r
e
,
th
e
r
e
c
ti
f
ie
d
li
ne
a
r
un
it
(
R
e
L
U
)
f
unc
ti
on
is
us
e
d,
th
is
in
c
r
e
a
s
e
s
th
e
s
tr
e
ngt
h
of
th
e
pha
s
e
s
th
a
t
ha
ve
in
c
r
e
a
s
e
d
va
lu
e
s
a
s
w
e
ll
a
s
r
e
s
tr
a
in
s
th
e
pha
s
e
s
th
a
t
ha
ve
le
s
s
e
r
va
lu
e
s
in
th
e
te
r
m
e
xpr
e
s
s
io
n
to
m
odi
f
y
th
e
e
xpr
e
s
s
io
n. A
ddi
ti
ona
ll
y,
th
e
s
p
a
ti
a
l
da
ta
a
s
w
e
ll
a
s
t
he
e
va
lu
a
ti
on
a
r
e
una
voi
da
bl
y
gr
ow
in
g
w
hi
le
us
in
g
a
tt
e
nt
io
n
m
ode
l
to
gr
a
s
p
th
e
de
pe
nd
e
nc
y
f
or
lo
ng
di
s
ta
n
c
e
.
H
e
nc
e
,
th
e
pool
in
g
m
e
c
ha
ni
s
m
is
us
e
d
to
le
s
s
e
n
th
e
s
iz
e
of
th
e
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
th
a
t
de
c
r
e
a
s
e
s
t
he
c
ount
of
pa
r
a
m
e
te
r
s
.
H
e
r
e
,
w
e
s
e
le
c
t
th
e
m
a
x
-
pool
t
o r
e
s
ul
t
in
e
nha
nc
e
d c
om
pr
e
s
s
io
n r
a
th
e
r
t
ha
n a
ve
r
a
ge
, w
hi
c
h i
s
gi
ve
n a
s
(
6)
. H
e
r
e
,
[
]
de
not
e
s
t
he
c
onc
e
a
le
d
e
xpr
e
s
s
io
ns
f
r
om
th
e
bi
di
r
e
c
ti
ona
l
L
S
T
M
a
s
w
e
ll
a
s
a
tt
e
nt
io
n
m
e
th
od,
a
s
w
e
ll
a
s
th
e
s
how
s
t
he
−
ℎ
phr
a
s
e
a
s
w
e
ll
a
s
t
he
nor
m
a
li
z
a
ti
on.
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
E
nhanc
in
g t
e
x
t
c
la
s
s
if
ic
at
io
n t
hr
ough nov
e
l
d
e
e
p l
e
ar
ni
ng
s
e
qu
e
nt
ia
l
at
te
nt
io
n f
us
io
n ar
c
hi
te
c
tu
r
e
(
Shi
lp
a)
4647
=
[
(
(
[
]
)
)
]
(
6)
3.3.
N
or
m
al
iz
at
io
n
T
he
bi
di
r
e
c
ti
ona
l
L
S
T
M
, a
tt
e
nt
io
n m
ode
l
a
s
w
e
ll
a
s
t
he
f
or
w
a
r
d l
a
ye
r
ha
ve
a
n i
nde
pe
nde
nt
s
tr
uc
tu
r
e
.
H
e
nc
e
,
w
e
ut
il
iz
e
d
a
r
e
s
id
ue
n
e
twor
k
to
om
it
th
e
gr
a
di
e
nt
a
bs
e
nc
e
,
w
hi
c
h
is
gi
ve
n
a
s
f
ol
lo
w
s
in
(
7)
.
I
n
th
e
gi
ve
n
(
7)
,
s
how
s
th
e
r
e
s
ul
t
of
a
f
te
r
th
e
f
or
w
a
r
d
ne
twor
k.
F
ur
th
e
r
,
w
e
s
e
le
c
t
th
e
e
xpr
e
s
s
io
n
a
s
it
ha
s
t
o be
s
ta
nd
a
r
di
z
e
d a
s
s
how
n i
n (
8)
.
=
+
(
7)
=
(
)
(
8)
3.4.
D
e
e
p
c
u
s
t
om
f
e
at
u
r
e
e
xt
r
ac
t
or
T
he
tr
a
di
ti
ona
l
m
ode
ls
w
oul
d
a
id
in
th
e
p
r
e
di
c
ti
on
of
c
la
s
s
e
s
th
a
t
a
r
e
r
e
s
ul
te
d
by
th
e
e
nc
ode
r
b
y
li
nki
ng
to
th
e
li
ne
a
r
la
ye
r
.
C
ons
id
e
r
in
g
th
e
pr
io
r
us
e
d
M
H
A
M
,
th
e
gr
a
s
pi
ng
of
f
e
a
tu
r
e
s
doe
s
not
s
how
to
be
e
f
f
ic
ie
nt
w
it
hout
pr
io
r
t
r
a
in
in
g
a
s
it
c
a
nnot
a
tt
a
in
e
f
f
e
c
ti
ve
r
e
s
ul
ts
,
m
a
in
ly
w
hi
le
th
e
te
r
m
s
ha
ve
a
n
in
c
r
e
a
s
e
d
le
ngt
h.
H
e
nc
e
,
th
e
ne
twor
k
i
s
de
e
p
e
ne
d
to
gr
a
s
p
gl
oba
l
d
a
ta
.
T
h
e
r
e
f
or
e
,
th
is
r
e
s
e
a
r
c
h
w
or
k
u
s
e
s
2D
c
onvolut
io
n
pr
oc
e
s
s
to
e
nha
n
c
e
th
e
e
xpr
e
s
s
io
n
of
th
e
e
m
be
dde
d
te
r
m
s
.
L
a
te
r
,
th
e
le
ngt
h
of
s
e
que
nc
e
is
r
e
duc
e
d
by
2
of
i
ts
in
it
ia
l
le
ngt
h
us
in
g
M
a
xP
ool
in
g.
T
he
pr
o
c
e
s
s
us
e
s
va
r
io
us
di
m
e
ns
io
ns
f
or
th
e
te
r
m
to
pr
oduc
e
a
tt
r
ib
ut
e
m
a
ps
a
s
gi
ve
n i
n (
9)
.
ℎ
=
(
.
:
+
−
1
+
)
(
9)
H
e
r
e
,
ℎ
de
not
e
th
e
:
+
−
1
f
e
a
tu
r
e
a
s
w
e
ll
a
s
th
e
a
c
ti
va
ti
on
f
unc
ti
on
r
e
s
pe
c
ti
ve
ly
.
de
not
e
s
th
e
f
il
te
r
a
nd
th
e
bi
a
s
is
s
how
n
a
s
.
F
ur
th
e
r
m
or
e
,
th
e
M
a
x
-
P
o
ol
in
g
pr
oc
e
s
s
f
or
f
e
a
tu
r
e
m
a
ps
to
c
a
pt
ur
e
th
e
hi
ghe
s
t
va
lu
e
f
e
a
tu
r
e
ℎ′
s
how
in
g
th
e
m
a
jo
r
f
e
a
tu
r
e
f
or
th
e
f
i
lt
e
r
,
in
w
hi
c
h
ℎ
′
=
{
ℎ
1
,
ℎ
2
,
…
.
.
,
ℎ
−
+
1
}
.
T
he
r
e
f
or
e
,
th
e
m
ode
l
is
s
how
n
w
it
h
two
ti
m
e
s
in
c
r
e
a
s
e
d
d
a
t
a
a
s
th
e
in
it
ia
l
phr
a
s
e
.
A
ddi
ti
ona
ll
y,
it
i
s
s
how
n
th
a
t
hi
gh
e
r
c
ount
of
a
tt
r
ib
ut
e
m
a
ps
le
a
d
s
to
twi
c
e
th
e
c
ount
of
out
put
c
ha
nne
ls
.
U
s
in
g
th
is
te
c
hni
que
w
oul
d
w
id
e
n
th
e
e
va
lu
a
ti
on,
a
lt
hough
it
w
oul
d
not
e
nha
nc
e
th
e
a
c
c
ur
a
c
y
dur
in
g
th
e
pr
oc
e
s
s
of
c
la
s
s
if
ic
a
ti
on.
T
he
r
e
f
or
e
,
th
is
is
r
e
s
ol
ve
d
by
th
e
f
e
a
tu
r
e
m
a
p
s
ha
vi
ng
a
n
a
ddi
ti
ona
l
pa
ddi
ng
pr
oc
e
s
s
,
a
f
te
r
w
hi
c
h
a
M
a
xP
ool
in
g
pr
oc
e
s
s
is
pe
r
f
or
m
e
d
to
r
e
duc
e
th
e
ti
m
e
of
e
va
lu
a
ti
on
by
ha
lf
f
o
r
e
ve
r
y
la
ye
r
of
c
onvolut
io
n.
A
r
e
s
id
ue
ne
twor
k
is
a
ls
o
u
s
e
d
to
s
to
p
th
e
a
b
s
e
nc
e
of
th
e
gr
a
di
e
nt
.
T
hi
s
is
e
xpr
e
s
s
e
d
a
s
gi
ve
n i
n (
10)
.
=
[
(
)
]
(
10)
=
[
(
(
)
)
]
+
(
11)
I
n
th
is
c
a
s
e
,
th
e
a
ddi
ti
on
s
ig
n
s
ig
ni
f
ie
s
th
e
r
e
s
id
ue
li
nk
r
e
pr
e
s
e
nt
s
th
e
f
in
a
l
c
onc
e
a
le
d
e
xpr
e
s
s
io
n
th
a
t
is
ut
il
iz
e
d
in
th
e
p
r
e
di
c
ti
on
of
th
e
pos
s
ib
le
c
la
s
s
out
put
us
in
g
S
of
tM
a
x.
C
ons
id
e
r
in
g
th
e
(
12)
,
s
how
s
th
e
−
ℎ
phr
a
s
e
.
T
he
w
e
ig
ht
is
de
not
e
d
us
in
g
a
nd
th
e
bi
a
s
i
s
gi
ve
n
us
in
g
.T
he
lo
s
s
is
c
a
lc
ul
a
te
d
a
nd
th
e
tr
a
in
in
g
a
tt
r
ib
ut
e
s
a
r
e
us
e
d
to
r
e
duc
e
th
e
lo
s
s
f
unc
ti
on.
T
he
tr
a
in
in
g
e
r
r
or
is
gi
ve
n
a
s
m
e
nt
io
ne
d
in
(
13)
.
T
he
r
e
a
l
ta
g
is
de
not
e
d
a
s
(
)
,
th
e
ta
g
th
a
t
is
pr
e
di
c
te
d
is
e
xpr
e
s
s
e
d
a
s
′
(
)
)
.
T
he
c
ount
of
phr
a
s
e
s
i
s
gi
ve
n
a
s
, t
he
t
ype
s
of
c
la
s
s
e
s
i
s
gi
ve
n a
s
.
′
(
)
=
(
|
)
′
(
)
=
(
+
)
(
12)
(
,
′
)
=
−
(
∑
=
1
∑
=
1
(
)
(
′
(
)
)
)
(
13)
4.
P
E
R
F
O
R
M
A
N
C
E
E
V
A
L
U
A
T
I
O
N
I
n
th
is
s
tu
dy,
th
e
b
e
nc
hm
a
r
k
da
ta
s
e
ts
,
na
m
e
ly
R
8,
M
R
,
a
nd
R
5
2
w
e
r
e
ut
il
iz
e
d
f
or
te
xt
c
la
s
s
if
ic
a
ti
on
ta
s
ks
.
T
h
e
s
e
da
ta
s
e
ts
e
n
c
om
pa
s
s
di
ve
r
s
e
dom
a
in
s
,
s
u
c
h
a
s
ne
w
s
a
r
ti
c
le
s
,
m
ovi
e
r
e
vi
e
w
s
,
a
nd
m
e
di
c
a
l
doc
um
e
nt
s
,
m
a
ki
ng
th
e
m
s
ui
ta
bl
e
f
or
e
va
lu
a
ti
ng
a
w
id
e
r
a
nge
of
te
xt
c
la
s
s
if
ic
a
ti
on
m
e
th
od
s
.
T
he
m
e
th
od
s
e
m
pl
oye
d
in
th
e
s
tu
dy
e
n
c
om
pa
s
s
a
s
p
e
c
tr
um
of
a
ppr
oa
c
he
s
,
f
r
om
tr
a
di
ti
ona
l
te
c
hni
que
s
li
ke
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
)
a
nd
f
a
s
tT
e
xt
to
a
dva
nc
e
d
d
e
e
p
le
a
r
ni
ng
m
ode
ls
li
ke
C
N
N
,
L
S
T
M
,
a
nd
B
E
R
T
.
T
he
e
va
lu
a
ti
on
pr
oc
e
s
s
in
vol
ve
s
m
e
a
s
ur
i
ng
th
e
pe
r
f
or
m
a
nc
e
of
th
e
s
e
m
e
th
ods
on
th
e
da
ta
s
e
ts
.
F
or
A
G
N
e
w
s
,
th
e
e
va
lu
a
ti
on
w
a
s
ba
s
e
d
on
M
a
c
r
o
-
F
1
s
c
or
e
s
a
nd
a
c
c
ur
a
c
y
th
e
r
e
s
ul
ts
a
r
e
s
how
n
in
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4642
-
4653
4648
th
e
f
or
m
of
gr
a
ph
a
nd
ta
bl
e
s
,
c
lo
s
e
ly
f
ol
lo
w
e
d
by
B
E
R
T
-
ba
s
e
d
m
ode
l
s
w
it
h
te
c
hni
que
s
li
k
e
li
ga
ti
on
in
de
pe
nde
nt
c
lo
ni
ng
(
L
I
C
)
a
nd
C
a
th
odol
um
in
e
s
c
e
nc
e
(
CL
)
. T
hi
s
e
xt
e
ns
iv
e
e
va
lu
a
ti
on f
r
a
m
e
w
or
k a
ll
ow
e
d f
o
r
a
r
obus
t
a
s
s
e
s
s
m
e
nt
of
te
xt
c
la
s
s
if
ic
a
ti
on
m
e
th
od
s
a
c
r
os
s
va
r
io
us
da
ta
s
e
t
s
,
s
how
c
a
s
in
g
th
e
na
tu
r
e
of
B
E
R
T
-
ba
s
e
d m
od
e
ls
i
n t
e
xt
c
la
s
s
if
ic
a
ti
on t
a
s
k
s
.
4.1.
D
at
as
e
t
d
e
t
ai
ls
T
he
be
nc
hm
a
r
k
da
ta
s
e
t
s
-
th
e
R
8
da
ta
s
e
t,
M
R
da
ta
s
e
t,
a
nd
R
52
da
ta
s
e
t
[
2]
a
r
e
us
e
d
in
th
e
s
tu
dy
to
c
onduc
t
th
e
in
ve
s
ti
ga
ti
ons
.
T
w
o
di
s
ti
nc
t
s
ub
s
e
ts
-
th
e
R
8
a
nd
R
52
da
ta
s
e
t
s
-
e
xt
r
a
c
te
d
f
r
om
th
e
R
e
ut
e
r
s
21,578
da
ta
s
e
ts
a
r
e
r
e
qui
r
e
d
to
be
u
s
e
d
in
or
de
r
to
c
om
pl
e
te
th
e
m
ul
ti
-
c
la
s
s
ta
s
k
s
. T
he
r
e
a
r
e
7,674
to
ta
l
doc
um
e
nt
s
in
th
e
c
ol
le
c
ti
on
known
a
s
R
8.
T
he
doc
um
e
nt
s
a
r
e
di
vi
de
d
in
to
two
s
e
pa
r
a
te
gr
oups
:
O
f
th
e
pa
pe
r
s
th
a
t
w
e
r
e
a
s
s
ig
ne
d,
2,189
w
e
r
e
e
xpr
e
s
s
ly
d
e
s
ig
na
te
d
f
or
te
s
ti
ng,
w
hi
le
th
e
r
e
s
t
5,485
w
e
r
e
a
s
s
ig
ne
d
f
or
tr
a
in
in
g.
E
ig
ht
s
e
pa
r
a
te
c
a
te
gor
ie
s
ha
ve
be
e
n a
ppl
ie
d t
o t
he
pa
pe
r
s
. T
w
o s
e
ts
o
f
t
he
da
ta
s
e
t
R
52 ha
ve
be
e
n c
r
e
a
te
d:
a
t
r
a
in
in
g
s
e
t
w
it
h
6,532
s
a
m
pl
e
s
a
nd
a
te
s
ti
ng
s
e
t
w
it
h
2,568
s
a
m
pl
e
s
.
5
2
c
a
te
gor
ie
s
a
r
e
pr
oduc
e
d
a
s
a
c
ons
e
que
nc
e
of
th
e
s
pl
it
. A
s
hor
t
te
xt
da
ta
s
e
t
us
e
d f
or
bi
na
r
y s
e
nt
im
e
nt
c
la
s
s
if
ic
a
ti
on i
s
c
a
ll
e
d t
he
M
R
d
a
ta
s
e
t.
T
he
pur
pos
e
of
th
e
s
of
twa
r
e
'
s
de
ve
lo
pm
e
nt
w
a
s
to
a
na
ly
s
e
m
ovi
e
r
e
vi
e
w
s
.
10,
662
r
e
vi
e
w
s
a
lt
oge
th
e
r
f
r
om
th
e
c
or
pus
,
w
hi
c
h
is
m
a
de
up
of
5,331
good
r
e
vi
e
w
s
a
nd
a
n
e
qua
l
num
be
r
of
ba
d
r
e
vi
e
w
s
.
T
a
bl
e
1
di
s
pl
a
y
s
th
e
s
um
m
a
r
y
of
th
e
da
ta
s
e
ts
.
T
a
bl
e
1
s
how
s
t
he
s
um
m
a
r
y of
t
he
da
ta
s
e
t
s
.
T
a
bl
e
1
.
S
um
m
a
r
y of
t
he
da
ta
s
e
t
s
D
a
t
a
s
e
t
D
oc
s
T
r
a
i
ni
ng
T
e
s
t
W
or
ds
N
ode
s
C
l
a
s
s
e
s
A
ve
r
a
ge
l
e
ngt
h
R8
7,674
5,485
2,189
7,688
15,362
8
65.72
MR
10,662
7,108
3,554
18,764
29,426
2
20.39
R
52
9,100
6,532
2,568
8,892
18,044
52
69.82
4.2.
A
G
N
E
WS
d
at
as
e
t
T
he
A
G
’
s
N
e
w
s
da
t
a
s
e
t
[
20]
w
hi
c
h
c
ont
a
in
s
127
,
600
s
a
m
pl
e
s
w
it
h
4
c
la
s
s
e
s
.
A
G
'
s
N
e
w
s
da
ta
s
e
t,
i
s
a
w
e
ll
-
known
a
nd
w
id
e
ly
us
e
d
te
xt
c
la
s
s
if
ic
a
ti
on
da
ta
s
e
t.
T
hi
s
da
ta
s
e
t
is
s
pe
c
if
ic
a
ll
y
de
s
ig
n
e
d
f
or
te
xt
c
a
te
gor
iz
a
ti
on
ta
s
k
s
,
m
a
ki
ng
it
va
lu
a
bl
e
f
or
tr
a
in
in
g
a
nd
e
va
lu
a
ti
ng
N
L
P
a
nd
m
a
c
hi
ne
le
a
r
ni
ng
m
ode
ls
.
T
he
da
ta
s
e
t
c
om
pr
is
e
s
ne
w
s
a
r
ti
c
le
s
c
ol
le
c
te
d
f
r
om
th
e
A
G
'
s
c
or
pu
s
,
c
ove
r
in
g
a
w
id
e
r
a
nge
of
to
pi
c
s
,
in
c
lu
di
ng
w
or
ld
ne
w
s
,
bus
in
e
s
s
,
s
por
ts
,
a
nd
s
c
ie
nc
e
.
E
a
c
h
ne
w
s
a
r
ti
c
le
i
s
a
s
s
oc
ia
te
d
w
it
h
a
c
or
r
e
s
ponding
c
a
te
gor
y
or
c
la
s
s
la
be
l,
m
a
ki
ng
it
s
ui
ta
bl
e
f
or
s
upe
r
vi
s
e
d
le
a
r
ni
ng
ta
s
ks
.
R
e
s
e
a
r
c
h
e
r
s
a
nd
da
ta
s
c
ie
nt
is
ts
c
om
m
onl
y
us
e
A
G
'
s
N
e
w
s
da
ta
s
e
t
to
de
ve
lo
p
a
nd
te
s
t
te
xt
c
l
a
s
s
if
ic
a
ti
on
a
lg
or
it
hm
s
,
s
e
nt
im
e
nt
a
na
ly
s
is
m
ode
l
s
,
a
nd
va
r
io
us
N
L
P
a
ppl
ic
a
ti
ons
due
t
o i
ts
r
e
a
l
-
w
or
ld
ne
w
s
c
ont
e
nt
a
nd dive
r
s
e
c
a
te
gor
ie
s
.
4.3.
R
e
s
u
lt
s
A
na
ly
s
in
g
th
e
p
e
r
f
or
m
a
nc
e
of
va
r
io
us
m
e
th
od
s
on
th
e
R
8
d
a
ta
s
e
t,
it
be
c
om
e
s
e
vi
de
nt
th
a
t
di
f
f
e
r
e
nt
a
ppr
oa
c
he
s
yi
e
ld
va
r
yi
ng
r
e
s
ul
ts
.
A
m
ong
th
e
s
e
m
e
th
od
s
,
th
e
pa
r
a
gr
a
ph
ve
c
to
r
s
w
it
h
a
di
s
tr
ib
ut
e
d
ba
g
of
w
or
ds
m
ode
l
(
PV
-
D
B
O
W
)
m
ode
l
a
c
hi
e
ve
s
a
m
ode
r
a
te
a
c
c
ur
a
c
y
of
85.87%
,
w
hi
le
p
a
r
a
gr
a
ph
ve
c
to
r
-
di
s
tr
ib
ut
e
d
m
e
m
or
y
(
PV
-
DM
)
la
gs
be
hi
nd
a
t
52.07%
,
in
di
c
a
ti
n
g
it
s
lo
w
e
r
e
f
f
ic
ie
nc
y.
f
a
s
tT
e
xt
s
ho
w
s
a
s
tr
ong
pe
r
f
or
m
a
nc
e
w
it
h
a
n
a
c
c
ur
a
c
y
of
96.13
%
,
w
hi
le
s
im
pl
e
w
o
r
d
-
e
m
be
ddi
ng
-
ba
s
e
d
m
ode
ls
(
S
W
E
M
)
a
nd
lo
w
e
ne
r
gy
a
va
il
a
bi
li
ty
in
m
a
le
s
(
L
E
A
M
)
a
ls
o
e
xhi
bi
t
a
ve
r
a
ge
r
e
s
ul
ts
of
95.32%
a
nd
93.31%
,
r
e
s
pe
c
ti
ve
ly
.
N
ot
a
bl
y,
te
xt
-
m
ul
ti
m
oda
l
gr
a
ph
ne
ur
a
l
ne
twor
k
(
M
G
N
N
),
e
xi
s
ti
ng
s
ys
te
m
(
ES
)
out
pe
r
f
or
m
s
th
e
ot
he
r
s
,
a
c
hi
e
vi
ng
a
n
a
c
c
ur
a
c
y
of
97.39%
.
T
r
a
di
ti
ona
l
m
e
th
ods
li
ke
T
F
-
I
D
F
+
lo
gi
s
ti
c
r
e
gr
e
s
s
io
n
(
LR
)
,
C
N
N
,
a
nd
L
S
T
M
yi
e
ld
a
c
c
ur
a
c
ie
s
of
93.74%
,
94.02%
,
a
nd
93.68%
,
r
e
s
p
e
c
ti
ve
ly
.
H
ow
e
ve
r
,
th
e
pos
t
s
c
r
ip
t
(
PS
)
m
ode
l
e
xhi
bi
ts
a
n
a
c
c
ur
a
c
y
of
98.13%
,
de
m
ons
tr
a
ti
ng
it
s
s
upe
r
io
r
c
a
pa
bi
li
ti
e
s
in
te
xt
c
la
s
s
if
ic
a
ti
on
on
th
e
R
8
da
ta
s
e
t.
T
a
bl
e
2
s
ho
w
s
th
e
c
om
pa
r
is
on
of
va
r
io
us
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
f
or
di
f
f
e
r
e
nt
da
ta
s
e
ts
.
F
ig
ur
e
2
s
how
s
t
he
c
om
pa
r
is
on on
R
8
d
a
ta
s
e
t.
T
a
bl
e
2
.
C
om
pa
r
is
on of
va
r
io
us
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
f
or
di
f
f
e
r
e
nt
da
ta
s
e
ts
M
e
t
hod
R8
R
52
MR
PV
-
D
B
O
W
[
21]
0.8587
0.7829
0.6109
PV
-
D
M
[
21]
0.5207
0.4492
0.5947
f
a
s
t
T
e
xt
[
22]
0.9613
0.9281
0.7514
S
W
E
M
[
23]
0.9532
0.9294
0.7665
L
E
A
M
[
24]
0.9331
0.9184
0.7695
T
e
xt
-
M
G
N
N
E
S
[
25]
0.9739
0.942
0.7746
TF
-
I
D
F
+ L
R
[
26]
0.9374
0.8695
0.7459
C
N
N
[
27]
0.9402
0.8537
0.7498
L
S
T
M
0.9368
0.8554
0.7506
PS
98.13
96.12
91.86
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
E
nhanc
in
g t
e
x
t
c
la
s
s
if
ic
at
io
n t
hr
ough nov
e
l
d
e
e
p l
e
ar
ni
ng
s
e
qu
e
nt
ia
l
at
te
nt
io
n f
us
io
n ar
c
hi
te
c
tu
r
e
(
Shi
lp
a)
4649
F
ig
ur
e
2
.
C
om
pa
r
is
on of
di
f
f
e
r
e
nt
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
w
it
h P
S
on R
8 D
a
ta
s
e
t
A
na
ly
s
in
g
th
e
pe
r
f
or
m
a
nc
e
of
va
r
io
us
m
e
th
ods
on
th
e
R
52
d
a
ta
s
e
t.
P
V
-
D
B
O
W
e
xhi
bi
ts
a
n
a
c
c
ur
a
c
y
of
78.29%
,
s
how
in
g
m
ode
r
a
te
pe
r
f
or
m
a
nc
e
,
w
h
e
r
e
a
s
P
V
-
D
M
tr
a
il
s
b
e
hi
nd
a
t
44.92%
,
in
di
c
a
ti
ng
li
m
it
e
d
e
f
f
e
c
ti
ve
ne
s
s
f
or
th
is
da
ta
s
e
t.
f
a
s
t
T
e
xt
a
nd
S
W
E
M
bot
h
p
e
r
f
or
m
w
e
ll
w
it
h
a
c
c
ur
a
c
ie
s
of
92.81%
a
nd
92.94%
,
r
e
s
pe
c
ti
ve
ly
.
L
E
A
M
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
91.84%
,
de
m
on
s
tr
a
ti
ng
a
s
tr
ong
c
la
s
s
if
ic
a
ti
on
c
a
p
a
bi
li
ty
.
T
he
T
e
xt
-
M
G
N
N
(
ES
)
m
e
th
od
pe
r
f
o
r
m
s
r
e
m
a
r
ka
bl
y
w
e
ll
,
w
it
h
a
n
a
c
c
ur
a
c
y
of
94.20%
.
I
n
c
ont
r
a
s
t,
t
r
a
di
ti
ona
l
m
e
th
ods
s
uc
h
a
s
T
F
-
I
D
F
+
L
R
,
C
N
N
,
a
nd
L
S
T
M
yi
e
ld
a
c
c
ur
a
c
ie
s
of
86.95%
,
85.37%
,
a
nd
85.54%
,
r
e
s
pe
c
ti
ve
ly
. N
ot
a
bl
y, t
he
P
S
m
ode
l,
a
c
hi
e
vi
ng a
n i
m
pr
e
s
s
iv
e
a
c
c
ur
a
c
y of
96.12%
. T
he
s
e
r
e
s
ul
ts
hi
ghl
ig
ht
t
he
va
r
yi
ng pe
r
f
or
m
a
nc
e
of
di
f
f
e
r
e
nt
m
ode
ls
, w
it
h P
S
e
m
e
r
gi
ng a
s
th
e
m
os
t
a
c
c
ur
a
te
c
hoi
c
e
f
or
t
e
xt
c
l
a
s
s
if
ic
a
ti
on
on t
he
R
52 da
ta
s
e
t
a
s
s
how
n i
n F
ig
ur
e
3.
A
na
ly
s
in
g
th
e
pe
r
f
or
m
a
nc
e
of
di
f
f
e
r
e
nt
m
e
th
ods
on
th
e
M
R
D
a
ta
s
e
t,
th
is
r
e
s
e
a
r
c
h
w
or
k
obs
e
r
ve
s
a
r
a
nge
of
a
c
c
ur
a
c
ie
s
.
P
V
-
D
B
O
W
a
nd
P
V
-
D
M
bot
h
yi
e
ld
s
r
e
la
ti
ve
ly
lo
w
a
c
c
ur
a
c
y,
w
it
h
s
c
or
e
s
of
61.09%
a
nd
59.47%
,
r
e
s
pe
c
ti
ve
ly
,
in
di
c
a
ti
ng
li
m
it
e
d
pe
r
f
or
m
a
nc
e
on
th
is
te
xt
c
la
s
s
if
ic
a
ti
on
ta
s
k.
f
a
s
tT
e
xt
a
nd
S
W
E
M
e
xhi
bi
t
m
ode
r
a
te
a
c
c
ur
a
c
y,
w
it
h
va
lu
e
s
of
75.14%
a
nd
76.65%
,
r
e
s
pe
c
ti
ve
ly
.
L
E
A
M
pe
r
f
or
m
s
s
li
ght
ly
be
tt
e
r
w
it
h
a
n
a
c
c
ur
a
c
y
of
76.95%
.
T
he
T
e
xt
-
M
G
N
N
(
ES
)
m
e
th
od
s
h
ow
s
c
om
m
e
nda
bl
e
pe
r
f
or
m
a
nc
e
,
a
c
hi
e
vi
ng
a
n
a
c
c
ur
a
c
y
of
77.46%
,
m
a
ki
ng
it
one
of
th
e
to
p
pe
r
f
or
m
e
r
s
in
th
is
da
t
a
s
e
t.
T
r
a
di
ti
ona
l
m
e
th
ods
li
ke
TF
-
I
D
F
+
L
R
,
C
N
N
,
a
nd
L
S
T
M
a
c
hi
e
v
e
a
c
c
ur
a
c
i
e
s
of
74.59%
,
74.98%
,
a
nd
75.06%
,
r
e
s
pe
c
ti
ve
ly
.
N
ot
a
bl
y,
th
e
P
S
m
ode
l
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
91.86%
.
T
he
s
e
r
e
s
ul
t
s
il
lu
s
tr
a
te
th
e
di
ve
r
s
it
y
in
m
ode
l
pe
r
f
or
m
a
nc
e
,
w
it
h P
S
e
m
e
r
gi
ng a
s
t
he
m
os
t
a
c
c
ur
a
te
c
hoi
c
e
f
or
t
e
xt
c
l
a
s
s
if
ic
a
ti
on on the
M
R
D
a
ta
s
e
t
a
s
s
how
n i
n F
ig
ur
e
4.
F
ig
ur
e
3
.
C
om
pa
r
is
on of
di
f
f
e
r
e
nt
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
w
it
h P
S
on R
52 D
a
ta
s
e
t
0
0
.1
0
.2
0
.3
0
.4
0
.5
0
.6
0
.7
0
.8
0
.9
1
A
c
c
u
r
a
c
y
v
a
l
u
e
M
et
h
o
d
o
l
o
gy
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
v
a
l
u
e
M
e
t
h
od
ol
og
y
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4642
-
4653
4650
F
ig
ur
e
4
.
C
om
pa
r
is
on of
di
f
f
e
r
e
nt
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
w
it
h P
S
on M
R
D
a
ta
s
e
t
T
a
bl
e
3
di
s
pl
a
ys
th
e
r
e
s
ul
t
s
f
or
a
c
c
ur
a
c
y
on
A
G
N
e
w
s
da
t
a
s
e
t
i
n
c
om
pa
r
is
on
w
it
h
va
r
io
us
m
ode
ls
in
a
c
la
s
s
if
ic
a
ti
on
ta
s
k,
B
E
R
T
+
L
I
C
+
C
L
is
th
e
s
e
c
ond
m
o
s
t
a
c
c
ur
a
te
m
ode
l
w
it
h
a
n
a
c
c
ur
a
c
y
of
91.15,
w
hi
c
h
is
not
a
bl
y l
ow
e
r
t
ha
n P
S
.
B
E
R
T
+
L
I
C
+
H
N
M
+
C
L
f
ol
lo
w
s
c
lo
s
e
ly
be
hi
nd, with a
n a
c
c
ur
a
c
y of
91.08. T
he
B
E
R
T
m
ode
l,
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
89.02,
s
how
c
a
s
in
g
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
pr
e
-
tr
a
in
e
d
la
ngua
ge
m
ode
ls
.
T
he
C
N
N
+
L
I
C
+
C
L
m
ode
l
a
ls
o
pe
r
f
or
m
s
w
e
ll
w
it
h
a
n
a
ve
r
a
g
e
a
c
c
ur
a
c
y
of
88.71.
G
e
ne
r
a
ll
y,
m
ode
ls
th
a
t
in
c
or
por
a
te
B
E
R
T
, pa
r
ti
c
ul
a
r
ly
w
it
h t
he
a
ddi
ti
on of
L
I
C
a
nd C
L
, de
li
ve
r
s
tr
ong a
c
c
ur
a
c
y, P
S
s
ta
nds
out
a
s
t
h
e
be
s
t
-
pe
r
f
or
m
in
g
m
ode
l
w
it
h
a
n
a
c
c
ur
a
c
y
of
95.97.
F
ig
ur
e
5
di
s
pl
a
ys
th
e
a
c
c
ur
a
c
y
c
om
pa
r
is
on
A
G
N
e
w
s
da
ta
s
e
t.
T
a
bl
e
3
.
C
om
pa
r
is
on on
A
G
N
e
w
s
D
a
ta
s
e
t
M
ode
l
A
c
c
ur
a
c
y
M
a
c
r
o
-
F1
C
N
N
[
26]
0.8797
0.8797
C
N
N
+L
I
C
0.8822
0.882
C
N
N
+C
L
0.8763
0.8757
C
N
N
+L
I
C
+
C
L
0.8871
0.8872
C
N
N
+L
I
C
+H
N
M
+C
L
0.8844
0.8848
B
E
R
T
[
28]
0.8902
0.8895
B
E
R
T
+L
I
C
0.8899
0.9022
B
E
R
T
+C
L
0.8893
0.8867
B
E
R
T
+L
I
C
+C
L
0.9115
0.9123
B
E
R
T
+L
I
C
+H
N
M
+C
L
0.9108
0.9127
PS
0.9597
0.9678
F
ig
ur
e
5
.
C
om
pa
r
is
on of
a
c
c
ur
a
c
y f
or
di
f
f
e
r
e
nt
s
ta
te
-
of
-
a
r
t
te
c
h
ni
que
s
w
it
h P
S
on AG
N
e
w
s
D
a
ta
s
e
t
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
Ac
c
u
r
a
c
y
v
a
l
u
e
Me
t
h
od
ol
og
y
0
.
8
2
0
.
8
4
0
.
8
6
0
.
8
8
0
.
9
0
.
9
2
0
.
9
4
0
.
9
6
0
.
9
8
A
ccu
r
a
cy v
a
l
u
e
Me
t
h
od
ol
og
y
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
E
nhanc
in
g t
e
x
t
c
la
s
s
if
ic
at
io
n t
hr
ough nov
e
l
d
e
e
p l
e
ar
ni
ng
s
e
qu
e
nt
ia
l
at
te
nt
io
n f
us
io
n ar
c
hi
te
c
tu
r
e
(
Shi
lp
a)
4651
T
he
M
a
c
r
o
-
F1
s
c
or
e
gr
a
ph
is
pl
ot
te
d
in
F
ig
ur
e
6
.
F
ol
lo
w
in
g
c
lo
s
e
ly
,
B
E
R
T
+
L
I
C
+
H
N
M
+
C
L
a
nd
B
E
R
T
+
L
I
C
+
C
L
bot
h
de
li
ve
r
hi
gh
M
a
c
r
o
-
F
1
s
c
or
e
s
of
91.27
a
nd
91.23,
r
e
s
pe
c
ti
ve
ly
,
m
a
ki
ng
th
e
m
s
ol
id
c
ont
e
nde
r
s
.
T
he
B
E
R
T
m
ode
l,
w
he
n
c
om
bi
ne
d
w
it
h
va
r
io
u
s
te
c
hni
que
s
li
ke
L
I
C
a
nd
C
L
,
c
ons
i
s
te
nt
ly
pe
r
f
or
m
s
w
e
ll
,
s
c
or
in
g
88.95,
90.22,
a
nd
88.67,
r
e
s
pe
c
ti
ve
ly
.
M
e
a
nw
hi
le
,
th
e
C
N
N
m
ode
ls
a
c
hi
e
v
e
a
ve
r
a
ge
s
c
or
e
s
,
w
it
h
C
N
N
+
L
I
C
+
C
L
a
nd
C
N
N
+
L
I
C
+
H
N
M
+
C
L
le
a
di
n
g
th
e
C
N
N
-
ba
s
e
d
m
ode
ls
w
it
h
s
c
or
e
s
of
88.72
a
nd
88.48,
r
e
s
p
e
c
ti
ve
ly
.
C
N
N
on
it
s
ow
n
a
c
hi
e
v
e
s
a
M
a
c
r
o
-
F
1
s
c
or
e
of
87.97,
w
hi
c
h
is
lo
w
e
r
th
a
n
th
e
to
p
-
pe
r
f
or
m
in
g
m
ode
ls
.
P
S
a
c
hi
e
ve
s
a
r
e
m
a
r
ka
bl
e
M
a
c
r
o
-
F
1
s
c
or
e
of
96.78,
s
ig
ni
f
yi
ng
it
s
r
obus
t
c
a
pa
bi
li
ty
f
or
t
e
xt
c
la
s
s
if
ic
a
ti
on t
a
s
ks
.
F
ig
ur
e
6
.
C
om
pa
r
is
on of
M
a
c
r
o
-
F
1 f
or
di
f
f
e
r
e
nt
s
ta
te
-
of
-
a
r
t
te
c
hni
que
s
w
it
h P
S
on AG
N
e
w
s
D
a
ta
s
e
t
4.4.
C
om
p
ar
at
iv
e
a
n
al
ys
is
I
n
th
e
e
va
lu
a
ti
on
of
m
ode
ls
a
c
r
os
s
f
our
da
ta
s
e
ts
-
A
G
N
e
w
s
,
R
8,
R
52,
a
nd
M
R
.
T
a
bl
e
4
pr
e
s
e
nt
s
a
c
om
pa
r
a
ti
ve
a
na
ly
s
is
of
t
w
o m
ode
ls
, E
S
a
nd P
S
, a
c
r
os
s
t
hr
e
e
di
s
ti
nc
t
da
ta
s
e
ts
:
R
8, R
52, a
nd
M
R
. T
he
m
e
tr
ic
s
us
e
d
f
or
e
va
lu
a
ti
on
in
di
c
a
te
th
e
pe
r
f
or
m
a
nc
e
of
th
e
s
e
m
od
e
ls
in
te
xt
c
la
s
s
if
ic
a
ti
on
ta
s
ks
.
N
ot
a
bl
y,
P
S
out
pe
r
f
or
m
s
E
S
in
a
ll
th
r
e
e
da
ta
s
e
ts
,
s
how
c
a
s
in
g
hi
ghe
r
a
c
c
u
r
a
c
y
in
R
8,
R
52,
a
nd
M
R
by
a
ppr
oxi
m
a
te
ly
0.76%
,
2.02%
,
a
nd
a
s
ig
ni
f
ic
a
nt
17.01%
,
r
e
s
pe
c
ti
ve
ly
.
T
hi
s
d
e
m
ons
tr
a
te
s
th
e
c
on
s
id
e
r
a
bl
e
e
f
f
e
c
ti
ve
n
e
s
s
of
th
e
P
S
m
ode
l,
pa
r
ti
c
ul
a
r
ly
in
th
e
M
R
da
ta
s
e
t,
w
he
r
e
it
s
pe
r
f
or
m
a
nc
e
im
pr
ove
m
e
nt
is
s
ubs
ta
nt
ia
l.
T
he
r
e
s
ul
ts
s
ugge
s
t
th
a
t
P
S
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
s
th
e
a
c
c
ur
a
c
y
of
te
xt
c
la
s
s
if
ic
a
ti
on
c
om
pa
r
e
d
to
E
S
a
c
r
os
s
th
e
s
e
di
ve
r
s
e
da
ta
s
e
ts
,
m
a
ki
ng
it
a
f
a
vor
a
bl
e
c
hoi
c
e
f
or
te
xt
c
la
s
s
if
ic
a
ti
on
ta
s
ks
.
T
h
e
im
pr
ove
m
e
nt
s
,
e
s
pe
c
ia
ll
y
in
th
e
M
R
da
ta
s
e
t,
hi
ghl
ig
ht
th
e
pot
e
nt
ia
l
of
a
dva
n
c
e
d
te
xt
c
la
s
s
if
ic
a
ti
on
m
ode
ls
to
a
c
hi
e
ve
hi
ghe
r
a
c
c
ur
a
c
y
in
ha
ndl
in
g
uns
tr
uc
tu
r
e
d t
e
xt
ua
l
da
ta
.
T
a
bl
e
4
.
C
om
pa
r
a
ti
ve
a
na
ly
s
is
D
a
t
a
s
e
t
ES
PS
I
m
pr
ovi
s
a
t
i
on
(%)
R8
0.9739
0.9813
0.756956
R
52
0.942
0.9612
2.01765
MR
0.7746
0.9186
17.0092
T
a
bl
e
5 pr
ovi
de
s
a
c
om
pa
r
a
ti
ve
a
na
ly
s
i
s
of
t
w
o m
ode
ls
, E
S
a
nd
P
S
, on the
A
G
N
e
w
s
da
ta
s
e
t
in
t
e
r
m
s
of
bot
h
a
c
c
ur
a
c
y
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
.
T
he
r
e
s
ul
ts
c
l
e
a
r
ly
de
m
ons
tr
a
te
th
a
t
th
e
P
S
m
ode
l
out
pe
r
f
or
m
s
E
S
in
bot
h
m
e
tr
ic
s
.
W
h
e
n
it
c
om
e
s
to
a
c
c
ur
a
c
y,
P
S
a
c
hi
e
ve
s
a
n
im
pr
e
s
s
iv
e
95.97%
,
w
hi
le
E
S
la
gs
be
hi
nd
w
it
h
91.08%
, i
ndi
c
a
ti
ng a
s
ubs
ta
nt
ia
l
5.23%
i
m
pr
ove
m
e
nt
. S
im
il
a
r
ly
, i
n t
e
r
m
s
of
M
a
c
r
o
-
F
1 s
c
or
e
s
, P
S
l
e
a
ds
w
it
h a
s
tr
ong
96.78%
,
s
ig
ni
f
ic
a
nt
ly
s
ur
pa
s
s
in
g
E
S
,
w
hi
c
h
s
c
or
e
s
91.
27%
.
T
hi
s
r
e
m
a
r
ka
bl
e
im
pr
ove
m
e
nt
of
5.86%
hi
ghl
ig
ht
s
t
he
e
f
f
ic
ie
nc
y of
t
he
P
S
m
ode
l
in
t
e
xt
c
la
s
s
if
ic
a
ti
on t
a
s
ks
w
it
hi
n t
he
A
G
N
e
w
s
da
t
a
s
e
t.
0
.8
2
0
.8
4
0
.8
6
0
.8
8
0
.9
0
.9
2
0
.9
4
0
.9
6
0
.9
8
M
a
c
r
o
F
1
_v
a
l
u
e
M
et
h
o
d
o
l
o
gy
Evaluation Warning : The document was created with Spire.PDF for Python.