I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
14
,
No.
6
,
De
c
e
mber
20
24
,
pp.
6689
~
6698
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
14
i
6
.
pp
6
689
-
6698
6689
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
A
sp
e
c
t
-
b
ase
d
se
n
t
im
e
n
t
-
an
al
y
si
s u
si
n
g t
op
ic
m
od
e
ll
in
g an
d
m
ac
h
i
n
e
-
l
e
ar
n
in
g
Radh
ik
a
Jin
e
n
d
r
a
Dhan
al
1
,
2
,
Vij
ay
Ra
m
G
h
or
p
ad
e
3
1
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g, D
. Y
. P
a
ti
l
C
ol
le
ge
of
E
ngi
ne
e
r
in
g a
nd T
e
c
hnol
ogy, S
hi
va
ji
U
ni
ve
r
s
it
y,
K
ol
ha
pur
, I
ndi
a
2
D
e
pa
r
tm
e
nt
of
T
e
c
hnol
ogy, S
hi
va
ji
U
ni
ve
r
s
it
y, K
ol
ha
pur
, I
ndi
a
3
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g, B
ha
r
a
ti
V
i
dya
pe
e
th
’
s
C
ol
le
ge
of
E
ngi
ne
e
r
in
g, S
hi
va
ji
U
ni
ve
r
s
it
y, K
ol
ha
p
ur
, I
n
di
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Apr
16,
2024
R
e
vis
e
d
Aug
6,
2024
Ac
c
e
pted
Aug
14,
2024
T
h
i
s
s
t
u
d
y
ad
d
res
s
es
t
h
e
cri
t
i
ca
l
n
eed
fo
r
a
n
accu
rat
e
as
p
ect
-
b
as
e
d
s
en
t
i
me
n
t
-
a
n
al
y
s
i
s
(A
BS
A
)
mo
d
e
l
t
o
u
n
d
er
s
t
a
n
d
s
e
n
t
i
m
en
t
s
effect
i
v
el
y
.
T
h
e
ex
i
s
t
i
n
g
A
BSA
mo
d
el
s
o
f
t
en
face
ch
a
l
l
e
n
g
e
s
i
n
accu
rat
e
l
y
ex
t
ract
i
n
g
as
p
ect
s
an
d
d
et
ermi
n
i
n
g
s
en
t
i
m
en
t
p
o
l
ar
i
t
y
fro
m
t
ex
t
u
a
l
d
at
a.
T
h
erefo
re,
w
e
p
ro
p
o
s
e
a
n
o
v
el
a
p
p
r
o
ach
l
ev
era
g
i
n
g
l
at
en
t
-
D
i
ri
c
h
l
e
t
-
al
l
o
ca
t
i
o
n
(L
D
A
)
f
o
r
as
p
ect
ex
t
rac
t
i
o
n
an
d
t
ran
s
fo
rmer
-
b
as
e
d
b
i
d
i
rec
t
i
o
n
a
l
-
en
c
o
d
er
-
rep
res
e
n
t
a
t
i
o
n
s
fro
m
t
ran
s
fo
rmer
s
(T
F
-
BE
R
T
)
fo
r
s
en
t
i
me
n
t
-
p
o
l
ari
t
y
ev
a
l
u
at
i
o
n
.
T
h
e
ex
p
er
i
men
t
s
w
ere
carri
e
d
o
u
t
o
n
Sem
E
v
a
l
2
0
1
4
l
a
p
t
o
p
a
n
d
re
s
t
a
u
r
an
t
d
at
a
s
et
s
.
A
l
s
o
,
a
mu
l
t
i
-
d
o
ma
i
n
d
at
a
s
et
w
a
s
g
e
n
erat
e
d
b
y
c
o
mb
i
n
i
n
g
Sem
E
v
a
l
2
0
1
4
,
A
maz
o
n
,
an
d
h
o
s
p
i
t
al
re
v
i
e
w
s
.
T
h
e
res
u
l
t
s
d
emo
n
s
t
ra
t
e
t
h
e
s
u
p
er
i
o
r
i
t
y
o
f
t
h
e
L
D
A
-
TF
-
B
E
RT
m
o
d
e
l
,
ach
i
e
v
i
n
g
8
2
.
1
9
%
accu
rac
y
an
d
7
9
.
5
2
%
Macro
-
F
1
s
co
re
fo
r
t
h
e
l
a
p
t
o
p
t
as
k
an
d
8
6
.
2
6
%
accu
rac
y
o
f
8
7
.
2
6
%
an
d
8
1
.
2
7
%
fo
r
Macro
-
F1
s
c
o
re
fo
r
t
h
e
res
t
au
ra
n
t
t
a
s
k
.
T
h
i
s
s
h
o
w
cas
e
s
t
h
e
mo
d
el
's
ro
b
u
s
t
n
e
s
s
an
d
effect
i
v
en
e
s
s
i
n
accu
ra
t
el
y
an
al
y
z
i
n
g
t
ex
t
u
a
l
d
at
a
an
d
ex
t
rac
t
i
n
g
mea
n
i
n
g
f
u
l
i
n
s
i
g
h
t
s
.
T
h
e
n
o
v
el
t
y
o
f
o
u
r
w
o
rk
l
i
es
i
n
c
o
mb
i
n
i
n
g
LD
A
an
d
T
F
-
BE
R
T
,
p
ro
v
i
d
i
n
g
a
co
mp
re
h
en
s
i
v
e
an
d
accu
ra
t
e
A
BSA
s
o
l
u
t
i
o
n
fo
r
v
ar
i
o
u
s
i
n
d
u
s
t
ri
e
s
,
t
h
ereb
y
co
n
t
ri
b
u
t
i
n
g
s
i
g
n
i
f
i
can
t
l
y
t
o
t
h
e
ad
v
a
n
cemen
t
o
f
s
e
n
t
i
men
t
an
al
y
s
i
s
t
ech
n
i
q
u
e
s
.
K
e
y
w
o
r
d
s
:
B
E
R
T
L
a
tent
Dir
ichle
t
a
ll
oc
a
ti
on
M
a
c
hine
lea
r
ning
S
e
mE
va
l
S
e
nti
ment
TF
-
B
E
R
T
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
a
dhika
J
inendr
a
Dha
na
l
De
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ienc
e
a
nd
E
nginee
r
ing
,
D.
Y.
P
a
ti
l
C
oll
e
ge
of
E
nginee
r
ing
a
nd
T
e
c
hnolo
gy,
S
hivaji
Unive
r
s
it
y
Kolha
pur
,
I
ndia
E
mail:
dha
na
ll
r
a
ddhe
ka
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
One
s
ubf
ield
of
na
tur
a
l
-
langua
ge
-
pr
oc
e
s
s
ing
(
N
L
P
)
known
a
s
a
s
pe
c
t
-
ba
s
e
d
s
e
nti
ment
-
a
na
ly
s
is
(
AB
S
A)
looks
f
or
pa
tt
e
r
ns
in
how
pe
ople
f
e
e
l
a
bout
c
e
r
tain
pa
r
ts
of
textua
l
c
ontent,
including
c
us
tom
e
r
-
f
e
e
dba
c
ks
,
s
oc
ial
-
m
e
dia
pos
ts
,
or
pr
oduc
t
-
r
e
view
s
[
1]
.
AB
S
A
di
f
f
e
r
s
f
r
om
c
onve
nti
ona
l
s
e
nti
ment
a
s
s
e
s
s
ment
methods
by
f
oc
us
ing
on
identif
ying
a
nd
e
va
luating
the
polar
it
y
of
s
e
nti
ment
(
ne
ut
r
a
l,
ne
ga
ti
ve
,
o
r
pos
it
ive)
towa
r
ds
s
pe
c
if
ic
a
s
pe
c
ts
a
long
with
e
nti
ti
e
s
dis
c
u
s
s
e
d
in
a
given
text
.
Unlike
c
onve
nti
ona
l
method
s
whic
h
of
f
e
r
a
ge
ne
r
a
l
s
e
nti
ment
r
a
ti
ng
to
the
e
nti
r
e
text
,
AB
S
A
a
tt
e
mpt
s
to
o
f
f
e
r
a
mo
r
e
gr
a
nular
e
va
lu
a
ti
on
by
e
xa
mi
ning
s
e
nti
ments
r
e
ga
r
ding
indi
v
idual
e
leme
nts
withi
n
the
text
[
2]
.
F
or
ins
tanc
e
,
in
a
r
e
s
taur
a
nt
-
r
e
view
,
AB
S
A
mi
ght
a
na
lyze
s
e
nti
ments
towa
r
ds
a
s
pe
c
ts
l
ike
f
ood
-
qua
li
ty,
a
mbi
a
nc
e
,
s
e
r
vice
,
a
nd
pr
icing
s
e
pa
r
a
tely
[
3]
.
AB
S
A
f
inds
a
ppli
c
a
ti
ons
a
c
r
os
s
va
r
ious
indus
tr
ies
whe
r
e
unde
r
s
tanding
c
us
tom
e
r
opini
ons
a
nd
s
e
nti
ments
towa
r
ds
s
pe
c
if
ic
f
e
a
tur
e
s
or
pr
oduc
ts
is
c
r
uc
ial.
I
n
the
r
e
tail
a
nd
e
-
c
omm
e
r
c
e
s
e
c
tor
,
AB
S
A
he
lps
bus
ines
s
e
s
a
na
lyz
e
c
us
tom
e
r
f
e
e
dba
c
k
on
pr
oduc
t
f
e
a
tur
e
s
,
pr
icing
,
s
hippi
ng,
a
nd
c
us
tom
e
r
s
e
r
vice
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
14
,
No.
6
,
De
c
e
mber
20
24
:
6689
-
6698
6690
im
pr
ove
c
us
tom
e
r
e
xpe
r
ienc
e
a
nd
make
da
ta
-
dr
iven
de
c
is
ions
[
4]
.
I
n
the
hos
pit
a
li
ty
indus
tr
y
,
AB
S
A
is
us
e
d
to
a
na
lyze
s
e
nti
ments
towa
r
d
hotel
a
meniti
e
s
,
r
oo
m
c
lea
nli
ne
s
s
,
s
taf
f
be
ha
vior
,
a
nd
dini
ng
opti
ons
t
o
e
nha
nc
e
gue
s
t
s
a
ti
s
f
a
c
ti
on
[
5]
.
I
t
i
s
a
ls
o
va
luable
in
s
e
c
to
r
s
li
ke
he
a
lt
hc
a
r
e
,
a
utom
oti
ve
,
f
inanc
e
,
a
nd
mor
e
,
whe
r
e
c
us
tom
e
r
f
e
e
dba
c
k
a
nd
s
e
nti
ment
a
na
ly
s
is
play
a
vit
a
l
r
ole
in
bus
ines
s
s
tr
a
tegie
s
a
nd
pr
oduc
t
de
v
e
lopm
e
nt
[
6]
.
T
he
ne
e
d
f
o
r
AB
S
A
a
r
is
e
s
f
r
om
the
li
mi
tations
of
tr
a
dit
ional
s
e
nti
ment
a
na
lys
is
[
7]
,
[
8]
whic
h
pr
ovides
a
ge
ne
r
a
li
z
e
d
s
e
nti
ment
s
c
or
e
without
c
o
ns
ider
ing
the
im
por
tanc
e
r
e
late
d
to
s
pe
c
if
ic
a
s
pe
c
ts
.
AB
S
A
e
na
bles
a
mor
e
gr
a
nular
unde
r
s
tanding
of
c
us
tom
e
r
opini
ons
,
a
ll
owing
bus
ines
s
e
s
to
pr
e
c
is
e
ly
identif
y
s
tr
e
ngths
,
we
a
kne
s
s
e
s
,
a
nd
a
r
e
a
s
f
o
r
im
pr
ove
me
nt.
How
e
ve
r
,
t
r
a
dit
ional
de
e
p
-
lea
r
ning
(
DL
)
[
9
]
,
[
10]
a
nd
mac
hine
-
lea
r
ning
(
M
L
)
[
11]
,
[
12]
models
us
e
d
in
AB
S
A
ha
ve
c
e
r
tain
dis
a
dva
ntag
e
s
.
M
or
e
ove
r
,
DL
models
in
AB
S
A
a
r
e
o
f
ten
tr
a
ined
on
s
pe
c
if
ic
da
tas
e
ts
a
nd
domains
,
making
them
les
s
a
da
ptable
to
ne
w
c
o
ntexts
or
indus
tr
ies
[
13]
.
T
he
y
may
a
ls
o
s
uf
f
e
r
f
r
om
b
ias
a
nd
lac
k
of
ge
ne
r
a
li
z
a
bil
it
y
,
a
s
they
f
oc
us
on
the
give
n
a
r
e
a
of
tr
a
ini
ng
a
nd
may
not
pe
r
f
or
m
we
ll
on
diver
s
e
da
t
a
s
e
ts
or
e
mer
ging
tr
e
nds
[
14]
.
Additi
ona
ll
y,
many
e
xis
ti
ng
models
in
AB
S
A
dir
e
c
tl
y
e
va
luate
the
ove
r
a
l
l
polar
it
y
of
r
e
view
s
without
c
ons
ider
ing
a
s
pe
c
t
-
leve
l
s
e
nti
ments
,
lea
ding
to
unr
e
li
a
ble
a
nd
bias
e
d
r
e
s
ult
s
[
15]
.
T
o
a
ddr
e
s
s
thes
e
c
ha
ll
e
nge
s
,
thi
s
wor
k
pr
e
s
e
nts
a
nove
l
a
ppr
oa
c
h,
i.
e
.
,
late
nt
-
Dir
ichle
t
-
a
ll
oc
a
ti
on
(
L
DA
)
f
o
r
a
s
pe
c
t
e
xtr
a
c
ti
on,
whic
h
c
a
n
identif
y
t
opics
or
a
s
pe
c
ts
withi
n
a
c
or
pus
of
r
e
view
s
.
L
DA
he
lps
in
identif
ying
r
e
late
d
wor
ds
to
thes
e
topi
c
s
,
wh
ich
a
r
e
c
ons
ider
e
d
a
s
a
s
pe
c
ts
f
or
s
e
nti
ment
a
na
lys
is
.
Additi
ona
ll
y,
a
dva
nc
e
d
models
li
ke
T
r
a
ns
f
or
mer
-
ba
s
e
d
bidi
r
e
c
ti
ona
l
-
e
nc
od
e
r
-
r
e
pr
e
s
e
ntatio
ns
f
r
om
tr
a
ns
f
or
mer
s
(
T
F
-
B
E
R
T
)
a
r
e
us
e
d
in
thi
s
wor
k
f
o
r
e
va
luating
s
e
nti
ment
polar
it
y,
both
a
t
the
ove
r
a
l
l
r
e
view
leve
l
a
nd
a
s
pe
c
t
leve
l.
B
y
leve
r
a
ging
L
DA
f
or
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
T
F
-
B
E
R
T
f
or
s
e
nti
ment
a
na
l
ys
is
,
the
pr
opos
e
d
a
ppr
oa
c
h
c
ombi
ne
s
a
s
pe
c
t
-
leve
l
s
e
nti
me
nt
a
na
lys
is
with
ove
r
a
ll
s
e
nti
ment
e
va
luation,
pr
o
vidi
ng
a
mor
e
c
ompr
e
he
ns
ive
unde
r
s
tanding
of
c
us
tom
e
r
s
e
nti
ments
.
T
his
a
ppr
oa
c
h
a
im
s
to
ove
r
c
ome
the
li
mi
tations
of
tr
a
dit
ional
AB
S
A
models
by
c
ons
ider
ing
diver
s
e
da
tas
e
t
s
,
e
xtr
a
c
ti
ng
a
s
pe
c
ts
dyna
mi
c
a
ll
y,
a
nd
e
va
luating
a
s
pe
c
t
-
leve
l
s
e
nti
ments
to
r
e
duc
e
bias
a
nd
im
p
r
o
ve
the
r
e
li
a
bil
it
y
of
s
e
nti
ment
a
na
lys
is
r
e
s
ult
s
in
va
r
ious
indus
tr
ies
.
T
he
c
ont
r
ibut
ions
of
thi
s
wor
k
c
a
n
be
s
umm
a
r
ize
d
a
s
f
oll
ows
:
−
I
ntr
oduc
e
d
a
nove
l
a
ppr
oa
c
h
ter
med
L
DA
-
TF
-
B
E
R
T
,
whic
h
c
ombi
ne
s
L
DA
f
or
a
s
pe
c
t
e
xtr
a
c
ti
o
n
a
nd
TF
-
B
E
R
T
f
or
s
e
nti
ment
a
na
lys
is
a
nd
polar
it
y
e
va
luation.
−
T
he
model's
pe
r
f
or
manc
e
wa
s
r
igo
r
ous
ly
e
va
luate
d
uti
li
z
ing
the
s
tanda
r
d
S
e
mE
va
l
-
2014
T
a
s
k
-
4
da
tas
e
t,
a
wide
ly
r
e
c
ognize
d
be
nc
hmar
k
f
o
r
s
e
nti
ment
a
na
lys
is
tas
ks
.
−
I
n
a
ddit
ion
to
the
S
e
mE
va
l
-
2014
T
a
s
k
-
4
da
tas
e
t,
thi
s
wor
k
ha
s
a
ls
o
be
e
n
e
va
luate
d
us
ing
diver
s
e
r
e
view
s
f
r
om
mu
lt
ipl
e
domains
,
including
hos
pit
a
l
a
nd
Am
a
z
on
r
e
view
s
.
−
T
he
r
e
s
ult
s
obtaine
d
f
r
om
the
L
DA
-
TF
-
B
E
R
T
m
ode
l
s
howc
a
s
e
be
tt
e
r
pe
r
f
or
manc
e
in
c
ompar
is
on
with
e
xis
ti
ng
s
e
nti
ment
a
na
lys
is
models
.
T
his
wor
k
is
s
tr
uc
tur
e
d
a
s
f
oll
ows
.
S
e
c
ti
on
2
de
lve
s
int
o
the
li
ter
a
tur
e
s
ur
ve
y,
p
r
ovidi
ng
a
n
ove
r
view
of
e
xis
ti
ng
r
e
s
e
a
r
c
h
a
nd
methodologi
e
s
in
s
e
nti
ment
-
a
na
lys
i
s
a
nd
AB
S
A.
I
n
s
e
c
ti
on
3,
the
L
DA
-
TF
-
B
E
R
T
model
is
thor
oughly
dis
c
us
s
e
d,
de
taili
ng
it
s
a
r
c
hit
e
c
tur
e
,
methodology,
a
nd
im
pleme
ntation
f
o
r
s
e
nti
ment
a
na
lys
is
a
nd
polar
it
y
e
va
luation
.
F
oll
owing
that,
s
e
c
ti
on
4
f
oc
us
e
s
on
e
va
luating
the
L
DA
-
TF
-
B
E
R
T
model
in
ter
ms
of
a
c
c
ur
a
c
y
a
nd
mac
r
o
-
F
-
s
c
or
e
,
pr
e
s
e
nti
ng
a
c
ompar
a
ti
ve
a
na
lys
is
with
c
ur
r
e
nt
wor
ks
f
or
highl
ight
ing
L
DA
-
TF
-
B
E
R
T
pe
r
f
or
manc
e
a
nd
a
dva
nc
e
ments
.
F
inally,
s
e
c
ti
on
5
e
nc
a
ps
ulate
s
the
c
onc
lus
ion
of
the
wo
r
k,
s
umm
a
r
izing
ke
y
f
indi
ngs
,
c
ont
r
ibut
i
ons
,
a
nd
potential
a
ve
nue
s
f
o
r
f
u
tur
e
r
e
s
e
a
r
c
h
in
th
e
a
r
e
a
of
s
e
nti
ment
-
a
na
ly
s
is
a
nd
AB
S
A.
2.
L
I
T
E
RA
T
UR
E
S
UR
VE
Y
T
he
li
ter
a
tu
r
e
s
ur
ve
y
de
lves
int
o
va
r
ious
methodo
logi
e
s
a
nd
f
indi
ngs
f
r
om
r
e
c
e
nt
s
tudi
e
s
in
AB
S
A
a
nd
s
e
nti
ment
-
a
na
lys
i
s
.
T
his
s
e
c
ti
on
c
ove
r
s
a
r
a
n
ge
of
a
ppr
oa
c
he
s
a
im
e
d
a
t
pr
e
dicting
s
e
nti
ment
p
olar
it
ies
a
nd
e
xtr
a
c
ti
ng
a
s
pe
c
t
-
s
e
nti
ment
pa
ir
s
withi
n
s
e
ntenc
e
s
or
doc
uments
.
In
[
16
]
,
the
objec
ti
ve
of
thi
s
s
tudy
wa
s
to
de
ve
lop
a
pr
e
dictive
model
f
o
r
s
e
nti
ment
-
polar
it
ies
a
s
s
oc
iate
d
with
va
r
ious
a
s
pe
c
ts
withi
n
doc
u
ments
or
s
e
ntenc
e
s
.
T
his
r
e
s
e
a
r
c
h
wa
s
mot
ivate
d
by
the
obs
e
r
va
ti
on
that
pr
e
vious
r
e
s
e
a
r
c
h
ha
d
ove
r
looked
th
e
c
r
uc
ial
f
a
c
tor
s
of
a
s
pe
c
t
s
e
nti
ment
-
polar
it
y
a
nd
loca
l
-
c
ontext
r
e
lations
hip.
T
o
ove
r
c
ome
thi
s
is
s
ue
,
thi
s
wor
k
pr
opos
e
d
a
loca
l
-
c
ontext
-
f
oc
us
(
L
C
F
)
a
ppr
oa
c
h
f
r
om
the
mul
ti
-
he
a
d
s
e
lf
-
a
tt
e
nti
on
(
M
HSA)
modul
e
pr
e
vious
ly
pr
e
s
e
nted
f
o
r
AB
S
A.
T
his
wo
r
k
u
ti
l
ize
d
two
laye
r
s
c
a
ll
e
d
c
ontext
-
f
e
a
tur
e
s
dyna
mi
c
-
we
ight
(
C
DW
)
a
nd
c
ontext
-
f
e
a
tur
e
s
dyna
mi
c
-
mas
k
(
C
DM
)
to
pr
ior
it
ize
loca
l
-
c
ontext
wor
ds
.
T
he
y
a
c
hieve
d
75.
08%
a
nd
74.
45%
a
c
c
ur
a
c
y
with
C
DM
a
nd
C
DW
,
r
e
s
pe
c
ti
ve
ly,
f
or
the
laptop
da
tas
e
t,
a
nd
80
.
98%
a
nd
80
.
89%
f
o
r
the
r
e
s
taur
a
nt
da
tas
e
t.
T
he
T
witt
e
r
da
tas
e
t
yielde
d
a
n
a
c
c
ur
a
c
y
of
71.
24%
.
F
ur
ther
,
P
ha
n
a
nd
Ogunbo
na
[
17]
,
e
xplor
e
d
s
e
ntenc
e
gr
a
mm
a
ti
c
a
l
-
a
s
pe
c
ts
a
nd
us
e
d
a
s
e
lf
-
a
tt
e
nti
on
a
ppr
oa
c
h
f
or
s
yntac
ti
c
a
l
-
lea
r
ning
a
nd
pr
opos
e
d
a
model
c
a
ll
e
d
L
C
F
on
s
yntax
a
s
pe
c
t
-
ba
s
e
d
-
c
las
s
if
ica
ti
on
(
L
C
F
S
-
ASC
)
.
T
he
y
c
ombi
ne
d
c
ontextua
li
z
e
d
e
mbeddings
(
i
.
e
.
,
R
oB
E
R
T
a
a
n
d
B
E
R
T
)
,
de
pe
nde
nc
y
-
ba
s
e
d,
a
nd
pa
r
t
-
of
-
s
pe
e
c
h
(
P
oS
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
s
pe
c
t
-
bas
e
d
s
e
nti
me
nt
-
analys
is
us
ing
topi
c
mode
ll
ing
and
…
(
R
adhika
J
inendr
a
Dhanal
)
6691
e
mbeddings
f
or
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
e
nha
nc
ing
pe
r
f
or
manc
e
.
T
he
L
C
F
S
-
ASC
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
80.
52
with
C
DW
a
nd
80.
34
with
C
DM
f
or
the
laptop
da
tas
e
t.
F
or
the
r
e
s
taur
a
nt
da
tas
e
t,
they
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
80.
31
with
C
DW
a
nd
80.
10
with
C
D
M
.
T
he
s
tudy
in
[
18
]
wa
s
c
e
nter
e
d
a
r
ound
the
e
x
tr
a
c
ti
on
o
f
a
s
pe
c
t
-
s
e
nti
ment
polar
it
y
-
pa
ir
s
.
T
he
a
ppr
oa
c
h
invol
ve
d
uti
li
z
ing
s
e
ntenc
e
inf
or
mation
,
e
dge
-
labe
l
s
,
a
nd
de
pe
nde
nc
y
-
tr
e
e
s
to
a
c
hieve
thi
s
goa
l.
F
o
r
thi
s
,
they
pr
opos
e
d
a
model
c
a
ll
e
d
ph
r
a
s
e
-
de
pe
nd
e
nc
y
gr
a
ph
-
a
tt
e
nti
on
-
ne
twor
k
(
P
D
-
R
GA
T
)
a
nd
c
onduc
ted
e
xpe
r
im
e
nts
with
Globa
l
-
Ve
c
tor
s
(
GloVE
)
a
nd
B
E
R
T
p
r
e
-
tr
a
ini
ng
models
.
P
D
-
R
GA
T
s
howe
d
be
tt
e
r
e
f
f
e
c
ti
ve
ne
s
s
c
ompar
a
ble
to
c
ur
r
e
nt
wor
ks
on
the
S
e
mE
va
l
-
2014
T
a
s
k
-
4
da
tas
e
t.
He
e
t
al.
[
19]
ma
in
f
oc
us
wa
s
to
unde
r
s
tand
the
r
e
lations
hip
be
twe
e
n
a
s
pe
c
t
s
e
nti
ment
-
polar
it
y
a
nd
global
-
c
ontext.
He
nc
e
,
pr
opos
e
d
a
mul
ti
li
ngua
l
lea
r
ning
model
us
ing
C
DM
a
nd
C
DW
c
a
ll
e
d
loca
l
a
nd
global
-
c
ontext
-
f
oc
us
C
DM
-
C
DW
(
L
GC
F
-
C
DM
-
C
DW
)
.
T
he
L
GC
F
-
C
D
M
-
C
DW
a
c
h
ieve
d
81.
29%
a
c
c
ur
a
c
y
f
or
the
laptop
da
tas
e
t
a
nd
85.
52%
a
c
c
ur
a
c
y
f
or
the
r
e
s
taur
a
nt
da
tas
e
t.
Yu
a
nd
Z
ha
ng
[
20]
f
oc
us
wa
s
to
de
tec
t
a
s
pe
c
t
-
ba
s
e
d
r
e
leva
nt
wor
ds
uti
li
z
ing
a
n
a
tt
e
nti
on
-
ba
s
e
d
a
ppr
oa
c
h.
T
his
wo
r
k
in
tr
oduc
e
d
a
n
a
s
pe
c
t
-
pos
it
ion
inf
or
mation
e
xtr
a
c
ti
on
a
ppr
oa
c
h
f
or
long
-
dis
tanc
e
de
pe
nde
nc
ies
(
s
e
nten
c
e
s
)
a
nd
p
r
opos
e
d
the
model
c
a
ll
e
d
mul
ti
-
we
ight
gr
a
ph
-
c
onvolut
ional
-
ne
twor
k
(
M
W
GC
N)
a
nd
a
c
hieve
d
a
c
c
ur
a
c
ies
of
78
.
8%
a
nd
83
.
82%
f
o
r
the
lap
top
a
nd
r
e
s
taur
a
nt
da
tas
e
ts
,
r
e
s
pe
c
ti
ve
ly.
M
a
e
t
al.
[
21
]
a
i
med
to
c
or
r
e
c
t
the
e
r
r
or
s
f
a
c
e
d
by
de
pe
nde
nc
y
-
pa
r
s
e
r
s
a
nd
to
unde
r
s
tand
the
c
ompl
e
x
s
e
ntenc
e
s
tr
uc
tur
e
s
,
he
nc
e
,
pr
e
s
e
nted
a
M
ult
i
-
GC
N
model.
T
he
M
ult
i
-
GC
N
model
a
c
hieve
d
a
c
c
ur
a
c
ies
of
798
.
8%
f
o
r
the
laptop
a
nd
83.
82
%
f
or
the
r
e
s
taur
a
nt
da
tas
e
t.
L
in
a
nd
J
oe
[
22]
pr
opos
e
d
a
mas
ke
d
a
tt
e
nti
on
mec
ha
nis
m
f
or
nois
e
r
e
duc
ti
on
a
nd
f
e
a
tur
e
e
nha
nc
e
ment
in
the
global
te
xt
c
a
ll
e
d
C
DM
-
C
DW
.
T
he
C
DM
-
C
DW
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
80
%
f
or
the
laptop
a
nd
80.
59
%
f
or
the
r
e
s
taur
a
nt
da
tas
e
ts
.
F
inally,
Z
ha
o
e
t
al.
[
23]
c
ons
ider
e
d
r
e
lati
ons
hips
a
mong
the
de
pe
nde
nc
y
-
e
dge
s
a
nd
node
s
f
or
a
s
pe
c
t
s
e
nti
ment
-
polar
it
y
de
tec
ti
on
a
nd
pr
opos
e
d
a
model
c
a
ll
e
d
s
tr
uc
tur
e
d
-
de
pe
nde
nc
y
tr
e
e
-
ba
s
e
d
GC
N
(
S
DT
GC
N)
a
ppr
oa
c
h.
T
he
S
DT
GC
N
a
c
hieve
d
a
c
c
ur
a
c
ies
of
7
8.
64%
f
or
the
laptop
a
nd
83.
32
%
f
o
r
the
r
e
s
taur
a
nt
da
tas
e
t.
W
hil
e
the
a
bove
s
tudi
e
s
pr
ovide
va
luable
ins
ight
s
a
nd
a
dva
nc
e
ments
in
AB
S
A,
it
i
s
e
s
s
e
nti
a
l
to
a
c
knowle
dge
c
e
r
tain
li
mi
tations
inher
e
nt
in
thes
e
methodologi
e
s
a
nd
f
ind
ings
.
S
ome
models
may
pe
r
f
or
m
we
ll
in
s
pe
c
if
ic
domains
li
ke
laptops
o
r
r
e
s
taur
a
nts
but
may
s
tr
uggle
whe
n
a
ppli
e
d
to
di
f
f
e
r
e
nt
do
mains
or
indus
tr
ies
.
Ada
pti
ng
thes
e
models
to
diver
s
e
d
omains
r
e
mains
a
c
ha
ll
e
nge
.
T
he
s
c
a
l
a
bil
it
y
of
models
,
e
s
pe
c
ially
thos
e
incor
por
a
ti
ng
a
tt
e
nti
on
mec
ha
nis
ms
or
gr
a
ph
-
ba
s
e
d
a
ppr
oa
c
he
s
,
may
be
li
mi
t
e
d
whe
n
de
a
li
ng
with
lar
ge
-
s
c
a
le
da
tas
e
t
s
or
r
e
a
l
-
ti
me
a
ppli
c
a
ti
ons
.
M
or
e
ove
r
,
the
a
dva
nc
e
d
models
dis
c
us
s
e
d
a
bove
r
e
quir
e
s
ubs
tantial
c
omput
a
ti
ona
l
r
e
s
our
c
e
s
a
nd
t
r
a
ini
ng
ti
me,
making
them
les
s
a
c
c
e
s
s
ibl
e
f
or
r
e
s
e
a
r
c
he
r
s
or
or
ga
niza
ti
ons
with
li
mi
ted
c
omput
ing
c
a
pa
bil
it
ies
.
Als
o,
li
ke
many
a
r
ti
f
icia
l
int
e
ll
igent
(
AI
)
s
ys
tems
,
a
s
pe
c
t
-
ba
s
e
d
s
e
nti
ment
c
las
s
if
ier
s
may
e
xhibi
t
bias
e
s
b
a
s
e
d
on
the
t
r
a
ini
ng
da
ta
,
potentially
lea
ding
to
unf
a
ir
o
r
bias
e
d
pr
e
dictions
,
e
s
pe
c
ially
c
onc
e
r
ning
s
e
ns
it
i
ve
topi
c
s
o
r
unde
r
r
e
pr
e
s
e
nted
g
r
oups
.
Give
n
t
he
r
a
pid
e
volut
ion
of
langua
ge
us
e
a
nd
e
mer
ging
tr
e
nds
,
models
de
ve
loped
ba
s
e
d
on
older
da
tas
e
ts
or
l
ingui
s
ti
c
pa
tt
e
r
ns
may
s
tr
uggle
to
a
da
pt
to
ne
w
li
nguis
ti
c
s
or
e
xpr
e
s
s
ions
.
He
nc
e
,
to
s
olve
a
ll
thes
e
pr
oblems
,
L
DA
-
TF
-
B
E
R
T
ha
s
be
e
n
pr
opos
e
d.
T
he
L
DA
a
p
pr
oa
c
h
he
lps
to
r
e
c
ognize
topi
c
s
a
long
with
r
e
lat
e
d
wor
ds
whic
h
he
lps
to
e
xtr
a
c
t
a
s
pe
c
ts
e
f
f
icie
ntl
y,
whe
r
e
a
s
the
T
F
-
B
E
R
T
he
lps
to
de
tec
t
the
polar
it
y
o
f
a
s
pe
c
ts
a
nd
s
e
ntenc
e
s
e
f
f
icie
ntl
y.
As
many
of
the
models
dis
c
us
s
e
d
a
bove
r
e
ly
he
a
vil
y
on
the
S
e
mE
va
l
-
2014
da
tas
e
t
f
or
e
va
luation,
he
nc
e
the
pr
opos
e
d
L
DA
-
TF
-
B
E
R
T
ha
s
a
ls
o
be
e
n
e
va
luate
d
uti
li
z
ing
the
S
e
m
E
va
l
-
2014
T
a
s
k
-
4
da
tas
e
t.
T
he
L
DA
-
TF
-
B
E
R
T
ha
s
be
e
n
dis
c
us
s
e
d
in
mor
e
de
pth
in
the
be
low
s
e
c
ti
on.
3.
M
ODE
L
I
n
F
igur
e
1,
the
L
DA
-
TF
-
B
E
R
T
model
a
r
c
hit
e
c
tur
e
is
pr
e
s
e
nted.
T
he
ini
ti
a
l
s
tep
of
the
L
DA
-
TF
-
B
E
R
T
model
invol
ve
s
e
xtr
a
c
ti
ng
r
e
view
text
f
r
o
m
the
c
or
pus
da
tas
e
t.
T
his
text
is
then
s
ubjec
ted
to
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
polar
it
y
e
xtr
a
c
ti
on
us
ing
de
dica
ted
models
,
i.
e
.
,
L
DA
a
nd
T
F
-
B
E
R
T
r
e
s
pe
c
ti
ve
ly.
T
h
e
a
s
pe
c
t
e
xtr
a
c
ti
on
model
is
r
e
s
pons
ibl
e
f
or
identif
ying
a
s
pe
c
ts
withi
n
the
text,
while
the
pola
r
it
y
e
xtr
a
c
ti
o
n
model
de
ter
mi
ne
s
the
polar
it
ies
of
a
s
pe
c
ts
a
nd
s
e
ntenc
e
s
.
Ulti
mate
ly,
by
c
ombi
n
ing
the
e
xtr
a
c
ted
a
s
pe
c
ts
,
s
e
ntenc
e
polar
it
ies
,
a
nd
a
s
pe
c
t
polar
it
ies
,
the
ove
r
a
ll
s
e
nti
m
e
nt
polar
it
y
o
f
the
s
e
ntenc
e
is
de
ter
mi
ne
d.
3.
1.
L
DA
m
od
e
l
f
or
as
p
e
c
t
e
xt
r
ac
t
ion
L
DA
ha
s
e
mer
ge
d
a
s
a
powe
r
f
ul
method
f
or
unc
ov
e
r
ing
hidden
pa
tt
e
r
ns
a
nd
s
tr
uc
tu
r
e
s
withi
n
r
e
view
da
tas
e
ts
(
de
noted
a
s
R
)
.
I
t
leve
r
a
ge
s
the
c
a
pa
bil
it
y
to
a
na
lyze
da
ta
a
nd
ge
ne
r
a
te
topi
c
-
ve
c
tor
s
by
modeling
li
ke
li
hood
dis
tr
ibut
ions
.
I
n
L
DA
,
a
s
de
picte
d
in
F
igur
e
2,
the
va
r
iable
s
a
r
e
de
f
ined
a
s
f
ol
lows
:
r
e
p
r
e
s
e
nts
the
pos
it
ion
of
a
wo
r
d
withi
n
a
s
pe
c
if
ic
r
e
view
,
,
a
nd
a
r
e
pa
r
a
mete
r
s
gove
r
ning
topi
c
s
a
nd
p
r
opor
ti
ons
,
r
e
s
pe
c
ti
ve
ly,
while
e
va
luate
s
the
mi
xtur
e
of
wor
ds
a
c
r
os
s
dif
f
e
r
e
nt
topi
c
s
.
K
s
igni
f
ies
ove
r
a
ll
topi
c
s
,
de
notes
the
dis
tr
ibut
ion
of
topi
c
s
f
or
e
ve
r
y
r
e
view
,
,
de
notes
the
matr
ix
e
nc
a
ps
ulating
r
e
view
s
a
nd
wo
r
ds
,
a
nd
,
r
e
pr
e
s
e
nts
the
li
ke
li
hood
dis
tr
ibut
ion
of
w
or
ds
.
F
u
r
ther
,
de
notes
the
ove
r
a
ll
r
e
view
s
,
a
nd
N
de
notes
the
ove
r
a
ll
wor
ds
pr
e
s
e
nt
in
r
e
view
s
.
I
n
th
e
L
DA
,
it
is
c
ons
ider
e
d
that
the
pr
ior
-
dis
tr
ibut
ion
o
f
r
e
view
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
14
,
No.
6
,
De
c
e
mber
20
24
:
6689
-
6698
6692
topi
c
s
f
oll
ows
a
Di
r
ichle
t
-
dis
tr
ibut
ion
[
24]
,
whi
c
h
is
a
f
unda
menta
l
a
s
pe
c
t
o
f
the
model
’
s
un
de
r
lyi
ng
pr
oba
bil
is
ti
c
f
r
a
mew
or
k
.
F
igur
e
1.
Ar
c
hit
e
c
tur
e
of
L
DA
-
TF
-
B
E
R
T
model
F
igur
e
2.
L
DA
model
I
n
L
DA
,
the
topi
c
dis
tr
ibut
ion
f
or
a
s
pe
c
if
ic
r
e
view
is
r
e
pr
e
s
e
nted
a
s
=
ℎ
(
⃗
)
,
whe
r
e
⃗
is
a
hype
r
-
pa
r
a
mete
r
ve
c
tor
with
dim
e
ns
ions
,
c
or
r
e
s
ponding
to
the
numbe
r
o
f
topi
c
s
.
T
he
a
s
s
ignm
e
nt
of
a
topi
c
to
the
ℎ
wor
d
in
r
e
view
R
,
de
noted
a
s
,
,
is
c
a
lcula
ted
us
ing
the
mul
ti
nomi
a
l
dis
tr
ibu
ti
on
a
s
,
=
(
)
.
I
t
is
to
be
noted
that
L
DA
c
ons
ider
s
the
pr
ior
di
s
tr
ibut
ion
of
wor
ds
withi
n
e
ve
r
y
topi
c
to
be
gove
r
ne
d
by
Dir
ichle
t
-
dis
tr
ibut
ion
a
s
dis
c
us
s
e
d
a
bove
.
S
pe
c
if
ica
ll
y
,
f
or
a
ny
topi
c
,
it
s
wor
d
-
dis
tr
i
buti
on
is
e
xpr
e
s
s
e
d
a
s
=
ℎ
(
η
⃗
⃗
)
,
whe
r
e
η
⃗
⃗
r
e
pr
e
s
e
nts
a
nother
hype
r
-
pa
r
a
mete
r
ve
c
tor
with
voc
a
bular
y
s
ize
dim
e
ns
ions
.
F
ur
ther
mor
e
,
the
obs
e
r
ve
d
wo
r
d
p
r
oba
bil
it
y
dis
tr
ibut
ion
f
or
e
a
c
h
wo
r
d
in
r
e
view
is
de
ter
mi
ne
d
a
s
=
(
)
,
whe
r
e
r
e
f
e
r
s
to
the
wor
d
d
is
tr
i
buti
on
a
s
s
oc
iate
d
with
the
topi
c
a
s
s
ignm
e
nt
,
f
or
the
given
wor
d
.
I
n
(
1
)
,
the
c
o
mpr
e
he
ns
ive
dis
tr
ibut
ion
e
nc
ompas
s
ing
a
ll
vis
ibl
e
a
nd
hidden
va
r
iable
s
withi
n
the
L
DA
model
c
a
n
be
e
f
f
icie
ntl
y
a
ppr
oxim
a
ted
thr
ough
Gibbs
s
a
mpl
ing,
a
te
c
hnique
c
omm
only
us
e
d
in
B
a
ye
s
ian
inf
e
r
e
nc
e
a
nd
p
r
oba
bil
is
ti
c
modeling
[
25]
.
T
his
pr
oc
e
s
s
a
ids
in
inf
e
r
r
ing
the
late
nt
topi
c
s
a
nd
wor
d
dis
tr
ibut
ions
withi
n
a
c
or
pus
of
doc
uments
,
e
na
bli
ng
the
e
xtr
a
c
ti
on
o
f
mea
ningf
ul
ins
ight
s
a
nd
pa
tt
e
r
ns
.
(
,
,
,
|
,
)
=
∏
(
|
=
1
)
(
|
)
(
|
)
(
|
)
(
1)
T
he
c
ohe
r
e
nc
e
s
c
or
e
f
or
e
a
c
h
topi
c
a
c
r
os
s
dif
f
e
r
e
nt
c
lus
ter
s
is
de
ter
mi
ne
d
by
e
va
luating
the
(
2)
[
26]
.
(
⃗
⃗
,
⃗
⃗
⃗
)
=
∑
∙
|
|
=
1
|
|
⃗
⃗
⃗
|
|
2
∙
|
|
⃗
⃗
⃗
|
|
2
(
2)
s
igni
f
ies
the
c
onf
ir
mation
mea
s
ur
e
,
while
r
e
pr
e
s
e
nts
the
pa
ir
of
ve
c
tor
s
a
nd
.
He
r
e
,
⃗
⃗
a
nd
⃗
⃗
⃗
de
note
the
c
os
ine
ve
c
tor
s
im
il
a
r
it
y.
F
oll
owing
the
a
s
s
e
s
s
ment
of
the
c
ohe
r
e
nc
e
s
c
or
e
us
ing
(
2)
,
the
topi
c
s
with
the
highes
t
li
ke
li
hood
a
r
e
c
hos
e
n.
T
his
s
e
lec
ti
on
pr
oc
e
s
s
is
f
a
c
il
it
a
ted
by
uti
li
z
ing
the
pyL
DA
vis
li
br
a
r
y,
whic
h
vis
ua
li
z
e
s
the
dis
ti
nc
t
topi
c
s
identif
ied.
F
r
om
th
e
dif
f
e
r
e
nt
topi
c
s
,
it
s
c
or
r
e
late
d
wor
ds
a
r
e
c
ons
ider
e
d
a
s
a
s
pe
c
ts
.
3.
2.
T
F
-
B
E
RT
m
od
e
l
f
or
p
olarit
y
c
las
s
if
icat
io
n
I
n
thi
s
s
e
c
ti
on,
a
hybr
id
joi
nt
-
e
xtr
a
c
ti
on
a
ppr
oa
c
h
us
ing
the
T
F
-
B
E
R
T
model
is
pr
e
s
e
nted
f
or
e
xtr
a
c
ti
ng
both
the
polar
i
ty
of
a
s
pe
c
ts
a
nd
r
e
view
s
.
I
n
thi
s
wor
k
,
the
B
E
R
T
ha
s
be
e
n
pr
e
-
tr
a
ined
a
nd
f
ine
-
tuned
uti
li
z
ing
a
tr
a
ns
f
or
mer
take
n
f
r
om
the
Huggi
ng
-
f
a
c
e
[
27]
.
C
ons
ider
the
da
tas
e
t
whic
h
r
e
pr
e
s
e
nt
s
a
s
e
t
of
r
e
view
s
,
whe
r
e
a
r
e
the
di
f
f
e
r
e
nt
s
e
ntenc
e
s
a
nd
r
e
pr
e
s
e
nts
the
tokens
(
wor
ds
)
.
T
his
c
a
n
be
r
e
pr
e
s
e
nted
a
s
(
3)
.
F
r
om
,
the
polar
i
ty
is
identi
f
ied
us
ing
a
tupl
e
s
e
t
,
a
s
pr
e
s
e
nted
in
(
4)
.
=
{
1
,
2
,
3
,
…
,
}
(
3)
=
{
(
,
,
)
}
|
|
(
4)
whe
r
e
r
e
pr
e
s
e
nts
the
a
s
pe
c
t
s
,
r
e
pr
e
s
e
nts
the
polar
it
y
of
a
s
pe
c
ts
,
r
e
pr
e
s
e
nts
the
polar
it
y
of
r
e
vie
ws
,
=
1
,
…
,
a
nd
|
|
r
e
pr
e
s
e
nts
how
many
s
e
ts
a
r
e
pr
e
s
e
nt
f
o
r
th
e
given
c
o
r
pus
da
tas
e
t.
M
or
e
ove
r
,
the
hyb
r
id
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
s
pe
c
t
-
bas
e
d
s
e
nti
me
nt
-
analys
is
us
ing
topi
c
mode
ll
ing
and
…
(
R
adhika
J
inendr
a
Dhanal
)
6693
joi
nt
-
e
xtr
a
c
ti
on
a
ppr
oa
c
h
uti
li
z
e
s
a
pos
it
ion
tagging
tec
hnique
f
or
de
f
in
ing
a
s
pe
c
ts
whic
h
a
r
e
pr
e
s
e
nt
f
r
om
the
s
tar
ti
ng
point
to
the
e
nding
point
of
a
s
e
ntenc
e
r
e
pr
e
s
e
nted
a
s
a
nd
.
S
im
il
a
r
ly,
the
pos
it
ion
taggi
ng
tec
hnique
is
a
ls
o
us
e
d
f
or
de
f
ini
ng
polar
it
y
p
r
e
s
e
nt
f
r
om
the
s
tar
ti
ng
point
to
the
e
nding
point
of
a
s
e
ntenc
e
r
e
pr
e
s
e
nted
a
s
a
nd
.
Als
o,
a
s
the
r
e
view
texts
a
r
e
e
mbedde
d,
thi
s
wor
k
uti
li
z
e
s
a
pos
it
ion
-
e
nc
od
ing
pr
oc
e
s
s
to
ge
t
the
pos
it
ion
of
a
s
pe
c
ts
a
nd
thei
r
a
s
s
oc
iate
d
polar
it
y.
Us
ing
the
B
E
R
T
,
the
tokens
a
r
e
f
ur
the
r
c
ontextua
li
z
e
d
a
nd
r
e
pr
e
s
e
nted
a
s
(
5)
.
=
{
1
,
2
,
…
,
}
(
5)
whe
r
e
,
r
e
pr
e
s
e
nts
a
s
e
t
of
ove
r
a
ll
tokens
a
n
d
r
e
pr
e
s
e
nts
indi
v
idual
tokens
.
F
ur
ther
,
f
or
t
he
c
las
s
if
ica
ti
on
of
tokens
,
thi
s
wor
k
uti
li
z
e
s
a
f
ine
-
tuni
ng
pr
oc
e
s
s
.
Dur
ing
f
ine
-
tuni
ng,
B
E
R
T
us
e
s
the
output
f
r
om
it
s
f
inal
laye
r
(
c
las
s
if
y
token)
a
s
input
to
a
S
of
tM
a
x
c
las
s
if
ier
.
T
he
S
of
tM
a
x
f
unc
ti
on
is
a
ppli
e
d
to
the
logi
ts
(
r
a
w
s
c
or
e
s
)
pr
oduc
e
d
by
the
c
las
s
if
ier
,
whic
h
c
onve
r
ts
thes
e
s
c
or
e
s
int
o
pr
oba
bil
it
ies
.
T
he
s
e
pr
oba
bil
it
ies
indi
c
a
te
the
li
ke
li
hood
of
e
a
c
h
c
las
s
(
e
.
g.
,
pos
it
ive,
ne
ga
ti
ve
,
ne
utr
a
l)
f
o
r
s
e
nti
ment
a
na
l
ys
is
,
a
nd
the
highes
t
pr
oba
bil
it
y
c
las
s
is
s
e
lec
ted
a
s
the
pr
e
dicte
d
s
e
nti
ment
labe
l.
T
he
pr
oba
bil
i
ty
is
de
f
ined
a
s
(
6)
.
=
(
+
)
(
6)
whe
r
e
,
is
the
we
ight
matr
ix
that
is
us
e
d
to
li
ne
a
r
l
y
tr
a
ns
f
or
m
the
c
ontextua
li
z
e
d
tokens
int
o
a
f
or
m
a
t
that
is
s
uit
a
ble
f
or
c
las
s
if
ica
ti
on
a
nd
is
a
bias
ve
c
tor
th
a
t
is
a
dde
d
to
the
li
ne
a
r
tr
a
ns
f
or
mation
output
to
s
hif
t
the
r
e
s
ult
a
nd
e
nha
nc
e
the
f
lexibil
it
y
of
the
B
E
R
T
.
Af
t
e
r
the
p
r
oc
e
s
s
of
c
las
s
if
ica
ti
on
of
tokens
,
the
ou
tcome
s
a
r
e
s
e
pa
r
a
ted
a
s
s
e
ntenc
e
polar
it
y
a
nd
a
s
p
e
c
t
polar
it
y
r
e
pr
e
s
e
nted
a
s
=
{
1
,
2
,
…
,
}
,
=
{
1
,
2
,
…
.
,
}
.
T
he
c
ompl
e
te
f
low
is
p
r
e
s
e
nted
in
F
igur
e
3
.
F
r
om
s
e
ntenc
e
polar
it
y
=
{
1
,
2
,
…
,
}
a
nd
a
s
pe
c
t
polar
it
y
=
{
1
,
2
,
…
,
}
,
the
ove
r
a
ll
s
e
nti
ment
polar
it
y
is
e
va
luate
d
us
ing
a
c
c
ur
a
c
y
a
nd
mac
r
o
F
-
s
c
or
e
a
s
(
7)
-
(
9)
.
=
+
+
+
+
(
7)
−
=
2
×
(
+
)
×
(
+
)
(
+
)
+
(
+
(
8)
−
=
−
(
9)
whe
r
e
,
r
e
pr
e
s
e
nts
,
r
e
pr
e
s
e
nts
,
r
e
pr
e
s
e
nts
a
nd
r
e
pr
e
s
e
nt
.
T
he
r
e
s
ult
s
f
or
the
L
DA
-
TF
-
B
E
R
T
model
a
r
e
e
va
luate
d
us
ing
(
7)
to
(
9)
.
F
ur
ther
,
in
the
ne
xt
s
e
c
ti
on,
the
c
ompl
e
te
r
e
s
ult
s
a
c
hieve
d
by
the
L
DA
-
TF
-
B
E
R
T
a
r
e
dis
c
us
s
e
d
in
de
tail.
F
igur
e
3.
T
F
-
B
E
R
T
model
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
14
,
No.
6
,
De
c
e
mber
20
24
:
6689
-
6698
6694
4.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
4.
1.
S
ys
t
e
m
r
e
q
u
ire
m
e
n
t
s
T
he
e
va
luation
of
the
L
DA
-
TF
-
B
E
R
T
model
wa
s
pe
r
f
or
med
ut
il
izing
a
n
I
ntel
i7
pr
oc
e
s
s
or
with
12
c
or
e
s
a
nd
16
GB
of
R
AM
,
ope
r
a
ti
ng
on
W
in
dows
11.
T
he
c
oding
f
or
the
model
wa
s
im
plem
e
nted
in
P
ython,
a
nd
the
e
xe
c
uti
on
wa
s
c
a
r
r
ied
out
wi
thi
n
t
he
Ana
c
onda
e
nvir
onment.
T
he
pa
r
a
mete
r
s
e
tt
ings
us
e
d
in
thi
s
wor
k
a
r
e
p
r
e
s
e
nted
in
T
a
ble
1.
T
a
ble
1.
P
a
r
a
mete
r
s
e
tt
ings
P
a
r
a
me
te
r
V
a
lu
e
D
r
opout
-
r
a
te
0.1
O
pt
im
iz
e
r
A
da
m
L
e
a
r
ni
ng r
a
te
2.00E
-
05
B
a
tc
h s
iz
e
32
M
a
x s
e
qu
e
nc
e
l
e
ngt
h
128
M
a
x e
poc
h
10
4.
2.
Dat
as
e
t
I
n
thi
s
wor
k,
hos
pit
a
l
da
tas
e
ts
we
r
e
c
oll
e
c
ted
f
r
om
[
28]
,
[
29
]
.
Als
o,
thi
s
wor
k
c
ons
ider
e
d
the
S
e
mE
va
l
2014
T
a
s
k
4
[
30
]
f
o
r
e
va
luation.
T
he
la
ptop
2014
a
nd
r
e
s
taur
a
nt
2014
a
r
e
two
s
ubtas
ks
withi
n
the
S
e
mE
va
l
s
e
nti
ment
a
na
lys
is
da
tas
e
t.
T
he
s
e
da
tas
e
ts
a
r
e
labe
led
a
s
ha
ving
dif
f
e
r
e
nt
a
s
pe
c
ts
a
nd
polar
it
y.
P
olar
it
y
include
s
s
e
nti
ments
,
i.
e
.
,
ne
ut
r
a
l,
ne
ga
ti
ve
,
or
pos
it
ive
a
s
s
oc
iate
d
with
thes
e
a
s
pe
c
ts
.
F
ur
ther
,
a
mul
ti
-
s
our
c
e
da
tas
e
t
ha
s
be
e
n
ge
ne
r
a
ted
f
or
thi
s
wor
k.
T
he
da
tas
e
t
c
ons
is
ts
of
inf
or
mation
c
oll
e
c
ted
f
r
o
m
diver
s
e
s
our
c
e
s
,
s
uc
h
a
s
Hos
pit
a
ls
[
28]
,
[
29]
,
S
e
mE
va
l
20
14
T
a
s
k
4
[
30]
,
a
nd
Ama
z
on
[
31
]
.
F
or
the
hos
pit
a
l
da
tas
e
t,
da
ta
e
xtr
a
c
ti
on
wa
s
c
onduc
ted
f
r
o
m
P
r
a
c
to
a
nd
M
ouths
hut
us
ing
a
we
b
-
s
c
r
a
pping
a
ppli
c
a
ti
on
pr
ogr
a
mm
ing
int
e
r
f
a
c
e
(
API
)
.
T
he
S
e
mE
va
l
2014
T
a
s
k
4
d
a
tas
e
t
include
d
e
xc
lus
ively
laptop
a
nd
r
e
s
taur
a
nt
da
ta.
Additi
ona
ll
y,
the
Ama
z
on
da
tas
e
t
c
ompr
is
e
d
in
f
or
mation
r
e
late
d
to
c
e
ll
phone
s
a
nd
e
lec
tr
onics
.
T
he
s
e
s
e
pa
r
a
te
da
tas
e
t
s
we
r
e
mer
ge
d
to
f
or
m
a
c
ompr
e
he
ns
ive
mul
ti
-
s
our
c
e
domain
da
tas
e
t.
4.
3.
As
p
e
c
t
p
olari
t
y
T
his
s
e
c
ti
on
de
lves
int
o
the
a
na
lys
is
of
a
s
pe
c
t
-
ba
s
e
d
s
e
nti
ment,
s
howc
a
s
ing
how
di
f
f
e
r
e
nt
a
s
pe
c
ts
withi
n
s
e
ntenc
e
s
c
a
n
be
a
s
s
oc
iate
d
with
va
r
ying
polar
it
y
s
c
or
e
s
,
a
nd
p
r
ovidi
ng
ins
ight
s
int
o
s
e
nti
ments
e
xpr
e
s
s
e
d
withi
n
text
da
ta.
I
n
thi
s
s
e
c
ti
on,
T
a
ble
2
is
dis
c
us
s
e
d,
whic
h
pr
e
s
e
nts
a
c
oll
e
c
ti
on
of
s
e
ntenc
e
s
a
long
with
their
c
or
r
e
s
ponding
a
s
pe
c
ts
a
nd
pola
r
it
ies
.
E
a
c
h
s
e
ntenc
e
is
a
na
lyze
d
to
identif
y
the
s
pe
c
if
ic
a
s
pe
c
t
be
ing
dis
c
u
s
s
e
d
a
nd
it
s
s
e
nti
ment
-
polar
it
y
r
e
late
d
to
that
a
s
pe
c
t.
F
or
ins
tanc
e
,
the
s
e
ntenc
e
“
B
oot
ti
me
is
s
upe
r
-
f
a
s
t
,
a
r
ound
a
nywhe
r
e
f
r
om
35
s
e
c
onds
to
1
mi
nute”
highl
ight
s
the
a
s
pe
c
t
of
s
pe
e
d,
whic
h
is
a
s
s
oc
iate
d
with
a
po
s
it
ive
polar
it
y
of
0.
C
onve
r
s
e
ly,
the
s
e
ntenc
e
“
tec
h
s
uppor
t
would
not
f
ix
the
pr
oblem
unles
s
I
bought
your
plan
f
o
r
$150
plus
”
f
oc
us
e
s
on
the
a
s
pe
c
t
of
the
pr
oblem,
with
a
ne
ga
ti
ve
polar
i
ty
of
-
1.
S
im
il
a
r
ly,
other
s
e
ntenc
e
s
s
uc
h
a
s
“
but
in
r
e
s
ume
thi
s
c
omput
e
r
r
oc
ks
!
”
e
mphas
ize
the
a
s
pe
c
t
of
the
c
omput
e
r
,
c
onve
ying
a
pos
it
ive
polar
it
y
o
f
1,
whi
le
“
S
e
tup
wa
s
e
a
s
y”
highl
ight
s
the
a
s
pe
c
t
of
s
e
tup
with
a
pos
it
ive
polar
it
y
.
L
a
s
tl
y,
the
s
e
ntenc
e
“
Did
no
t
e
njoy
the
ne
w
W
indows
8
a
nd
touchs
c
r
e
e
n
f
unc
ti
o
ns
”
doe
s
not
s
pe
c
if
y
a
pa
r
ti
c
ular
a
s
pe
c
t
a
nd
is
labe
led
a
s
“
N
one
”
with
a
ne
ga
ti
ve
polar
it
y
o
f
0
.
T
a
ble
2.
S
e
ntenc
e
s
a
nd
their
r
e
late
d
a
s
pe
c
ts
a
nd
polar
it
y
S
e
nt
e
nc
e
A
s
pe
c
t
P
ol
a
r
it
y
B
oot
t
im
e
i
s
s
upe
r
-
f
a
s
t
, a
r
ound a
nyw
he
r
e
f
r
om 35 s
e
c
ond
s
t
o 1
mi
nut
e
.
F
a
s
t
0
te
c
h s
uppor
t
w
oul
d not f
ix
t
he
pr
obl
e
m unl
e
s
s
I
bought your
pl
a
n f
or
$150 plus
.
P
r
obl
e
m
-
1
but
i
n r
e
s
ume
t
hi
s
c
omput
e
r
r
oc
ks
!
C
omput
e
r
1
T
he
s
e
tu
p w
a
s
e
a
s
y.
S
e
tu
p
1
D
id
not
e
nj
oy t
he
ne
w
W
in
dow
s
8 a
nd t
ou
c
hs
c
r
e
e
n f
unc
ti
on
s
.
N
one
0
4.
4.
S
e
n
t
e
n
c
e
p
olarit
y
T
his
s
e
c
ti
on
pr
ovides
a
de
tailed
br
e
a
kdown
o
f
s
e
nti
ment
a
na
lys
is
,
s
howc
a
s
ing
how
pr
oba
bil
it
ies
in
dif
f
e
r
e
nt
s
e
nti
ment
c
a
tegor
ies
c
ontr
ibut
e
to
the
o
ve
r
a
ll
s
e
nti
ment
polar
it
y
of
e
a
c
h
s
e
ntenc
e
.
I
n
thi
s
s
e
c
ti
on,
T
a
ble
3
is
dis
c
us
s
e
d,
whic
h
pr
ovides
a
n
a
na
lys
i
s
of
s
e
nti
ment
polar
it
y
f
o
r
a
s
e
t
of
s
e
ntenc
e
s
.
T
he
table
include
s
c
olum
ns
f
or
ne
ga
ti
ve
,
ne
utr
a
l
,
a
nd
pos
it
i
ve
pr
oba
bil
it
ies
,
a
long
with
the
ove
r
a
ll
s
e
nti
ment
polar
it
y
s
c
or
e
.
E
a
c
h
s
e
ntenc
e
is
e
va
luate
d
to
de
ter
mi
ne
the
pr
oba
bil
it
y
dis
tr
ibut
ion
a
c
r
os
s
thes
e
s
e
nti
ment
c
a
t
e
gor
ies
,
indi
c
a
ti
ng
the
de
gr
e
e
of
ne
ga
ti
vi
ty,
ne
ut
r
a
li
ty,
a
nd
pos
it
ivi
ty
e
xpr
e
s
s
e
d
in
the
text.
F
or
ins
tanc
e
,
the
s
e
ntenc
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
s
pe
c
t
-
bas
e
d
s
e
nti
me
nt
-
analys
is
us
ing
topi
c
mode
ll
ing
and
…
(
R
adhika
J
inendr
a
Dhanal
)
6695
“
B
oot
ti
me
is
s
upe
r
-
f
a
s
t
,
a
r
ound
a
nywhe
r
e
f
r
o
m
35
s
e
c
onds
to
1
mi
nute”
ha
s
a
r
e
latively
ba
lanc
e
d
dis
tr
ibut
ion
a
mong
the
thr
e
e
c
a
tegor
ies
,
r
e
s
ult
ing
in
a
ne
ut
r
a
l
s
e
nti
ment
polar
it
y
s
c
or
e
o
f
0
.
C
onve
r
s
e
ly,
the
s
e
ntenc
e
“
tec
h
s
uppor
t
would
not
f
ix
the
p
r
oble
m
unles
s
I
bought
you
r
plan
f
or
$150
plus
”
ha
s
a
high
pr
oba
bil
it
y
in
the
ne
ga
ti
ve
c
a
tegor
y,
lea
ding
to
a
n
e
ga
ti
ve
s
e
nti
ment
polar
it
y
s
c
or
e
of
-
1.
S
im
il
a
r
ly
,
s
e
ntenc
e
s
li
ke
“
but
in
r
e
s
ume
thi
s
c
omput
e
r
r
oc
ks
!
”
a
nd
“
S
e
t
up
wa
s
e
a
s
y”
e
xhibi
t
s
tr
ong
pos
it
ive
pr
oba
bil
it
ies
,
r
e
s
ult
ing
in
pos
it
ive
s
e
nti
ment
polar
it
y
s
c
or
e
s
of
1
f
or
both
s
e
ntenc
e
s
.
On
the
other
ha
nd,
the
s
e
ntenc
e
“
Did
not
e
njoy
the
ne
w
W
indows
8
a
nd
touchs
c
r
e
e
n
f
unc
ti
ons
”
s
hows
a
pr
e
domi
na
nt
p
r
oba
bil
it
y
in
the
n
e
ga
ti
ve
c
a
tegor
y,
c
ontr
ibut
ing
to
a
ne
ga
ti
ve
s
e
nti
ment
pola
r
it
y
s
c
or
e
of
-
1.
T
a
ble
3.
S
e
ntenc
e
s
polar
it
y
s
c
or
e
s
S
e
nt
e
nc
e
N
e
ga
ti
ve
N
e
ut
r
a
l
P
os
it
iv
e
P
ol
a
r
it
y
B
oot
t
im
e
i
s
s
upe
r
f
a
s
t,
a
r
ound a
nyw
h
e
r
e
f
r
om 35 s
e
c
onds
t
o 1
mi
nut
e
.
0.0313
0.5616
0.4069
0
te
c
h s
uppor
t
w
oul
d not f
ix
t
he
pr
obl
e
m unl
e
s
s
I
bought your
pl
a
n f
or
$150 plus
.
0.78885
0.19955
0.0118
-
1
but
i
n r
e
s
ume
t
hi
s
c
omput
e
r
r
oc
ks
!
0.005
0.0518
0.9427
1
S
e
tu
p w
a
s
e
a
s
y.
0.021
0.3667
0.6119
1
D
id
not
e
nj
oy t
he
ne
w
W
in
dow
s
8 a
nd t
ou
c
hs
c
r
e
e
n f
unc
ti
on
s
.
0.927
0.06666
0.0060
-
1
4.
5.
Over
all
s
e
n
t
i
m
e
n
t
p
olarit
y
T
he
r
e
s
ult
s
f
r
om
the
S
e
mE
va
l
l
a
ptop
tas
k
e
va
luation
r
e
ve
a
l
int
e
r
e
s
ti
ng
ins
ight
s
int
o
the
pe
r
f
o
r
manc
e
of
va
r
ious
models
in
AB
S
A
a
s
pr
e
s
e
nted
in
F
igu
r
e
4.
S
tar
t
ing
with
the
R
GA
T
model
,
it
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
78.
02%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
74,
indi
c
a
ti
ng
a
de
c
e
nt
pe
r
f
or
manc
e
in
identi
f
ying
a
s
pe
c
ts
a
nd
s
e
nti
ments
withi
n
laptop
-
r
e
late
d
r
e
view
s
.
M
oving
on
to
M
ult
i
GC
N,
it
s
howe
d
a
s
li
ght
im
pr
ove
ment
with
a
n
a
c
c
ur
a
c
y
of
78.
8%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
74.
97
,
s
howc
a
s
ing
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
c
a
ptur
ing
s
yntac
ti
c
a
nd
s
e
mantic
inf
or
mation
f
o
r
s
e
nti
ment
a
na
lys
is
tas
ks
.
T
he
M
W
GC
N
model
de
mons
tr
a
ted
f
ur
ther
im
pr
ove
men
t
with
a
n
a
c
c
ur
a
c
y
of
79
.
78%
a
nd
a
r
e
mar
ka
ble
M
a
c
r
o
-
F
1
s
c
or
e
of
76
.
68,
highl
ight
ing
it
s
c
a
pa
bil
it
y
t
o
e
xtr
a
c
t
c
ontextua
l
f
e
a
tur
e
s
a
nd
a
s
pe
c
t
-
r
e
late
d
s
e
mantics
a
c
c
ur
a
tely.
T
he
C
DM
-
C
DW
model
s
howc
a
s
e
d
c
o
mpetit
ive
pe
r
f
or
manc
e
with
a
n
a
c
c
ur
a
c
y
of
80%
a
nd
a
n
im
pr
e
s
s
ive
M
a
c
r
o
-
F
1
s
c
or
e
of
79.
81,
indi
c
a
ti
ng
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
leve
r
a
ging
c
ontext
f
e
a
tur
e
s
a
nd
d
yna
mi
c
we
ight
ing
f
or
s
e
nti
ment
a
na
lys
is
tas
k
s
.
F
inally,
the
L
DA
-
TF
-
B
E
R
T
model
e
mer
ge
d
a
s
the
top
pe
r
f
or
mer
with
a
n
a
c
c
ur
a
c
y
of
82.
19%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
79.
52,
s
howc
a
s
ing
it
s
s
upe
r
ior
it
y
in
u
ti
li
z
ing
L
DA
f
or
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
T
F
-
B
E
R
T
f
o
r
s
e
nti
ment
polar
it
y
e
va
luation.
F
igur
e
4.
P
e
r
f
or
manc
e
e
va
luation
us
ing
laptop
-
2014
da
tas
e
t
T
he
r
e
s
ult
s
f
r
om
the
S
e
m
E
va
l
r
e
s
taur
a
nt
tas
k
e
va
luation
pr
ovide
va
luable
ins
ight
s
int
o
the
pe
r
f
or
manc
e
of
dif
f
e
r
e
nt
models
in
a
s
pe
c
t
-
ba
s
e
d
s
e
nti
ment
a
na
lys
is
withi
n
the
r
e
s
taur
a
nt
domain
a
s
pr
e
s
e
nted
in
F
igu
r
e
5
.
B
e
ginni
ng
with
the
R
GA
T
model,
i
t
a
c
hieve
d
a
notable
a
c
c
ur
a
c
y
of
83
.
55%
a
nd
a
r
e
s
pe
c
table
M
a
c
r
o
-
F
1
s
c
or
e
of
75.
99
,
s
howc
a
s
ing
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
identif
ying
a
s
pe
c
ts
a
nd
s
e
nti
ments
withi
n
r
e
s
taur
a
nt
r
e
view
s
.
T
he
M
ult
iGC
N
model
de
mons
tr
a
ted
a
s
li
ght
im
pr
ove
ment
with
a
n
a
c
c
ur
a
c
y
o
f
83.
82
%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
77.
1,
indi
c
a
ti
ng
i
ts
c
a
pa
bil
it
y
t
o
c
a
ptur
e
s
yntac
ti
c
a
nd
s
e
mantic
inf
o
r
mation
f
o
r
s
e
nti
ment
a
na
lys
is
ta
s
ks
in
the
r
e
s
taur
a
nt
domain.
M
ovin
g
on
to
the
M
W
GC
N
model,
it
s
howc
a
s
e
d
s
i
gnif
ica
nt
im
pr
ove
ment
with
a
n
im
pr
e
s
s
ive
a
c
c
ur
a
c
y
of
86.
36%
a
nd
a
n
outs
tanding
M
a
c
r
o
-
F
1
s
c
or
e
o
f
80
.
54,
highl
ight
ing
i
ts
a
bil
it
y
to
e
xtr
a
c
t
c
ontextua
l
f
e
a
tur
e
s
a
nd
a
s
pe
c
t
-
r
e
late
d
s
e
mantics
a
c
c
ur
a
tely,
ther
e
by
e
nha
nc
ing
s
e
nti
ment
a
na
lys
i
s
pe
r
f
or
manc
e
s
ubs
tantially.
T
he
C
DM
-
C
DW
model,
a
lt
hough
s
h
owing
a
78.02
78.8
79.78
80
82.19
74
74.97
76.68
79.81
79.52
65
70
75
80
85
R
GAT
M
u
lt
i
GCN
M
W
GC
N
C
D
M
-
C
D
W
L
D
A
-
T
F
B
E
R
T
(
%
)
M
o
d
e
ls
P
e
r
for
m
an
c
e
Eval
u
ati
on
of Lap
top
-
2014
D
atas
e
t
L
ap
t
op
A
c
c
u
r
ac
y
L
ap
t
op
M
ac
r
o-
F
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
14
,
No.
6
,
De
c
e
mber
20
24
:
6689
-
6698
6696
c
ompetit
ive
a
c
c
ur
a
c
y
of
80.
5
%
,
f
e
ll
s
li
ghtl
y
be
hin
d
with
a
M
a
c
r
o
-
F
1
s
c
or
e
of
80.
3
,
s
ugge
s
ti
ng
that
while
it
e
f
f
e
c
ti
ve
ly
leve
r
a
ge
d
c
ontext
f
e
a
tur
e
s
a
nd
dyna
m
ic
we
ight
ing,
the
r
e
is
r
oom
f
o
r
im
pr
ove
ment
in
c
a
ptur
ing
s
e
nti
ments
withi
n
r
e
s
taur
a
nt
r
e
view
s
.
F
inally,
the
L
DA
-
TF
-
B
E
R
T
model
e
mer
ge
d
a
s
the
top
pe
r
f
o
r
mer
with
a
n
e
xc
e
pti
ona
l
a
c
c
ur
a
c
y
of
87
.
26%
a
nd
a
n
im
pr
e
s
s
ive
M
a
c
r
o
-
F
1
s
c
or
e
of
81
.
27,
s
howc
a
s
ing
it
s
s
upe
r
ior
it
y
in
L
DA
f
or
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
T
F
-
B
E
R
T
f
or
s
e
nti
ment
polar
it
y
e
va
luation.
T
his
de
mons
tr
a
tes
the
e
f
f
e
c
ti
ve
ne
s
s
of
c
ombi
ning
topi
c
modeling
a
nd
DL
tec
hniques
f
or
r
obus
t
AB
S
A
in
the
r
e
s
taur
a
nt
domain,
lea
ding
to
a
n
e
nha
nc
e
d
unde
r
s
tanding
o
f
c
us
tom
e
r
s
e
nti
ments
a
nd
pr
e
f
e
r
e
nc
e
s
.
F
igur
e
5.
P
e
r
f
or
manc
e
e
va
luation
us
ing
laptop
-
2014
da
tas
e
t
F
ur
ther
,
T
a
ble
4
pr
e
s
e
nts
a
c
ompar
a
ti
ve
s
tudy
of
v
a
r
ious
models
in
ter
ms
o
f
a
c
c
ur
a
c
y
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
f
or
bo
th
the
laptop
a
nd
r
e
s
taur
a
nt
tas
ks
,
a
long
with
a
mul
t
i
-
c
a
tegor
y
e
va
luation.
E
a
c
h
model's
pe
r
f
or
manc
e
is
a
na
lyze
d
a
c
r
os
s
dif
f
e
r
e
nt
ye
a
r
s
,
pr
ovidi
ng
ins
ight
s
int
o
their
e
f
f
e
c
ti
ve
ne
s
s
in
AB
S
A.
S
tar
ti
ng
with
the
L
C
F
-
B
E
R
T
-
C
D
M
a
nd
L
C
F
-
B
E
R
T
-
C
DW
models
f
r
om
2019,
they
s
howc
a
s
e
c
ompetit
ive
pe
r
f
or
manc
e
with
a
c
c
ur
a
c
y
r
a
nging
f
r
om
79.
91
%
to
80.
05
%
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
r
a
nging
f
r
o
m
7
9.
69%
to
79.
85%
in
the
l
a
ptop
tas
k.
S
im
il
a
r
ly
,
in
the
r
e
s
taur
a
nt
tas
k,
their
a
c
c
ur
a
c
y
r
a
nge
s
f
r
o
m
80.
13
%
to
80
.
2%
,
with
M
a
c
r
o
-
F
1
s
c
or
e
s
be
twe
e
n
79.
97%
a
nd
80.
03%
.
T
h
e
s
e
models
e
s
tablis
h
a
s
oli
d
f
ounda
ti
on
f
or
AB
S
A.
M
oving
on
to
the
R
GA
T
model
f
r
o
m
2022,
it
e
xhibi
ts
de
c
e
nt
a
c
c
ur
a
c
y
in
the
laptop
tas
k
(
78.
02%
)
but
pe
r
f
or
ms
be
tt
e
r
in
the
R
e
s
taur
a
nt
tas
k
with
a
n
a
c
c
ur
a
c
y
of
83.
55%
.
How
e
ve
r
,
it
s
M
a
c
r
o
-
F
1
s
c
or
e
s
a
r
e
c
ompar
a
ti
v
e
ly
lowe
r
,
s
ugge
s
ti
ng
r
oom
f
or
im
pr
ove
ment
in
c
a
ptur
ing
s
e
nti
ments
.
T
he
L
GC
F
-
C
DM
-
C
DW
model
f
r
om
20
19
s
hows
pr
omi
s
ing
r
e
s
ult
s
,
e
s
pe
c
ially
in
the
r
e
s
taur
a
nt
tas
k,
with
a
c
c
ur
a
c
y
r
e
a
c
hing
85
.
52%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
79.
85%
.
T
his
indi
c
a
tes
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
ha
nd
li
ng
AB
S
A
in
r
e
s
taur
a
nt
-
r
e
late
d
r
e
view
s
.
T
he
p
r
opos
e
d
L
DA
-
TF
-
B
E
R
T
model
outper
f
or
ms
mos
t
of
th
e
pr
e
vious
models
,
s
howc
a
s
ing
s
upe
r
ior
a
c
c
ur
a
c
y
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
a
c
r
os
s
a
ll
tas
ks
.
I
n
the
L
a
ptop
tas
k,
it
a
c
hieve
s
a
n
a
c
c
ur
a
c
y
of
82.
19%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
79.
52,
while
in
the
r
e
s
taur
a
nt
tas
k,
it
a
c
hieve
s
a
n
a
c
c
ur
a
c
y
of
86.
26%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
80.
27.
T
he
mul
ti
-
domain
e
va
luation
f
u
r
ther
highl
ight
s
it
s
r
ob
us
tnes
s
,
with
a
n
a
c
c
ur
a
c
y
of
87
.
14%
a
nd
a
M
a
c
r
o
-
F
1
s
c
or
e
of
85
.
15.
Ove
r
a
ll
,
the
c
ompar
a
ti
ve
s
tudy
unde
r
s
c
or
e
s
the
a
dva
nc
e
ments
in
a
s
pe
c
t
-
ba
s
e
d
s
e
nti
ment
a
na
lys
is
models
,
with
ne
we
r
models
l
ike
L
DA
-
TF
-
B
E
R
T
d
e
mons
tr
a
ti
ng
s
upe
r
ior
pe
r
f
or
manc
e
in
a
c
c
ur
a
tely
d
e
tec
ti
ng
a
s
pe
c
ts
a
nd
s
e
nti
ments
in
r
e
view
s
,
thus
c
ontr
ibut
ing
s
igni
f
ica
ntl
y
to
mar
ke
t
ins
ight
s
a
nd
c
us
tom
e
r
unde
r
s
tanding.
T
a
ble
4.
C
ompar
a
ti
ve
s
tudy
M
ode
ls
L
a
pt
op
R
e
s
ta
ur
a
nt
M
ul
ti
-
D
oma
in
A
c
c
ur
a
c
y
M
a
c
r
o
-
F1
A
c
c
ur
a
c
y
M
a
c
r
o
-
F1
A
c
c
ur
a
c
y
M
a
c
r
o
-
F1
L
C
F
-
B
E
R
T
-
C
D
M
, 2019,
[
16]
79.91
79.69
80.2
80.03
-
-
L
C
F
-
B
E
R
T
-
C
D
W
, 2019,
[
16]
80.05
79.85
80.13
79.97
-
-
L
C
F
S
-
B
E
R
T
-
C
D
M
, 2020,
[
17]
79.88
79.56
80.34
80.19
-
-
L
C
F
S
-
B
E
R
T
-
C
D
W
, 2020,
[
17]
79.93
79.66
80.41
80.25
-
-
R
G
A
T
, 2022,
[
18]
78.02
74
83.55
75.99
-
-
L
G
C
F
-
C
D
M
-
C
D
W
[
19]
81.29
78.86
85.52
79.85
-
-
M
W
G
C
N
, 2023
[
20]
79.78
76.68
86.36
80.54
-
-
M
ul
ti
-
G
C
N
, 2023,
[
21]
78.8
74.97
83.82
77.1
-
-
C
D
M
-
C
D
W
, 2023
[
22]
80
79.81
80.59
80.39
-
-
S
D
T
G
C
N
, 2024
[
23]
78.64
75.50
83.32
76.13
-
-
L
D
A
-
TF
-
B
E
R
T
(
P
r
opos
e
d)
82.19
79.52
87.26
81.27
87.14
85.15
83.55
83.82
86.36
80.5
87.26
75.99
77.1
80.54
80.3
81.27
70
75
80
85
90
R
GAT
M
u
lt
i
GCN
M
W
GC
N
C
D
M
-
C
D
W
L
D
A
-
T
F
B
E
R
T
(
%
)
M
o
d
e
ls
P
e
r
for
m
an
c
e
Eval
u
ati
on
of R
e
s
tau
r
an
t
-
2014
D
atas
e
t
R
e
s
t
au
r
an
t
A
c
c
u
r
ac
y
R
e
s
t
au
r
an
t
M
ac
r
o-
F
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
s
pe
c
t
-
bas
e
d
s
e
nti
me
nt
-
analys
is
us
ing
topi
c
mode
ll
ing
and
…
(
R
adhika
J
inendr
a
Dhanal
)
6697
5.
CONC
L
USI
ON
I
n
c
onc
lus
ion,
thi
s
c
ompr
e
he
ns
ive
wo
r
k
ha
s
de
l
ve
d
de
e
p
int
o
the
r
e
a
lm
o
f
AB
S
A.
T
hr
ough
a
s
ys
tema
ti
c
a
ppr
oa
c
h,
va
r
ious
models
a
nd
metho
dologi
e
s
we
r
e
e
va
luate
d
a
nd
c
ompar
e
d
a
c
r
os
s
dif
f
e
r
e
nt
domains
,
including
the
laptop
a
nd
r
e
s
taur
a
nt
tas
ks
,
to
dis
c
e
r
n
their
e
f
f
ica
c
y
in
e
xtr
a
c
ti
ng
a
s
pe
c
ts
a
nd
de
ter
mi
ning
s
e
nti
ment
pola
r
it
y
f
r
om
textua
l
da
ta.
T
he
s
tudy
be
ga
n
with
a
n
e
xplor
a
ti
on
of
e
xis
ti
ng
l
i
ter
a
tur
e
,
highl
ight
ing
the
im
por
tanc
e
of
AB
S
A
in
unde
r
s
t
a
nding
c
us
tom
e
r
s
e
nti
ments
,
mar
ke
t
tr
e
nds
,
a
nd
pr
oduc
t
e
va
luations
.
T
he
ne
e
d
f
or
a
c
c
ur
a
te
a
nd
e
f
f
icie
nt
A
B
S
A
models
wa
s
e
mphas
iz
e
d,
e
s
pe
c
ially
in
indus
tr
ies
s
uc
h
a
s
e
-
c
omm
e
r
c
e
,
hos
pit
a
li
ty,
a
nd
tec
hnology,
whe
r
e
c
us
tom
e
r
f
e
e
dba
c
k
plays
a
pivot
a
l
r
ole.
T
he
methodology
de
tailed
the
L
DA
-
TF
-
B
E
R
T
model
a
nd
e
va
luation.
Dive
r
s
e
da
tas
e
ts
f
r
om
s
our
c
e
s
li
ke
S
e
mE
va
l,
Ama
z
on,
a
nd
hos
pit
a
l
r
e
view
s
we
r
e
uti
li
z
e
d
to
c
r
e
a
te
a
c
ompr
e
he
ns
ive
mul
ti
-
s
our
c
e
domain
da
tas
e
t,
e
ns
ur
ing
the
r
obus
tnes
s
a
nd
ge
ne
r
a
li
z
a
ti
on
o
f
the
models
.
T
he
e
xpe
r
im
e
ntal
r
e
s
ult
s
s
howc
a
s
e
d
the
pe
r
f
o
r
manc
e
o
f
va
r
ious
models
,
including
R
GA
T
,
M
ult
i
-
GC
N,
M
W
GC
N,
C
DM
-
C
DW
,
a
nd
the
pr
opos
e
d
L
DA
-
TF
-
B
E
R
T
model.
T
he
e
va
luation
metr
ics
s
uc
h
a
s
a
c
c
ur
a
c
y
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
we
r
e
a
na
lyze
d
a
c
r
os
s
dif
f
e
r
e
nt
ta
s
ks
a
nd
domains
,
pr
ovid
ing
va
luable
ins
ight
s
int
o
e
a
c
h
model's
s
tr
e
ngths
a
nd
li
mi
tations
.
Nota
bly
,
the
L
DA
-
TF
-
B
E
R
T
model
e
mer
ge
d
a
s
a
top
pe
r
f
o
r
mer
,
de
mons
tr
a
ti
ng
s
upe
r
ior
a
c
c
ur
a
c
y
a
nd
M
a
c
r
o
-
F
1
s
c
or
e
s
in
both
the
L
a
ptop
a
nd
R
e
s
taur
a
nt
tas
ks
.
I
ts
a
bil
it
y
to
leve
r
a
ge
L
DA
f
or
a
s
pe
c
t
e
xtr
a
c
ti
on
a
nd
T
F
-
B
E
R
T
f
o
r
s
e
nti
ment
polar
it
y
e
va
luation
pr
ove
d
highl
y
e
f
f
e
c
ti
ve
,
s
how
c
a
s
ing
it
s
potential
f
or
r
e
a
l
-
wor
ld
a
ppli
c
a
ti
ons
in
mar
ke
t
a
na
lys
is
a
nd
c
us
tom
e
r
f
e
e
dba
c
k
int
e
r
pr
e
tation
.
I
n
f
utu
r
e
wo
r
k,
the
model
c
a
n
be
e
va
luate
d
us
ing
dif
f
e
r
e
nt
da
tas
e
ts
to
va
li
da
te
it
s
pe
r
f
or
manc
e
.
RE
F
E
RE
NC
E
S
[
1]
G
.
K
ont
ona
ts
io
s
e
t
al
.
,
“
F
A
B
S
A
:
a
n
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
da
ta
s
e
t
of
us
e
r
r
e
vi
e
w
s
,”
N
e
ur
o
c
om
put
in
g
,
vol
.
562,
D
e
c
.
2023, doi:
10.1016/j
.ne
uc
om.2023.126867.
[
2]
S
.
D
a
tt
a
a
nd
S
.
C
ha
kr
a
ba
r
ti
,
“
A
s
pe
c
t
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
f
or
de
mone
ti
z
a
ti
on
twe
e
ts
by
opt
im
iz
e
d
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
tw
or
k
us
in
g f
ir
e
f
ly
-
or
ie
nt
e
d mul
ti
-
ve
r
s
e
opt
im
iz
e
r
,”
Sādhanā
, vol
. 46, no. 2, J
un. 2021, doi:
10.1007/s
12046
-
021
-
01608
-
1.
[
3]
G
.
K
a
ur
a
nd
A
.
S
ha
r
ma
,
“
H
A
S
:
hybr
id
a
na
ly
s
is
of
s
e
nt
im
e
nt
s
f
or
th
e
pe
r
s
pe
c
ti
ve
of
c
us
to
me
r
r
e
vi
e
w
s
umm
a
r
iz
a
ti
on,”
J
ou
r
n
al
of
A
m
bi
e
nt
I
nt
e
ll
ig
e
nc
e
and Hum
ani
z
e
d C
om
put
in
g
, vol
. 14, no. 9,
pp. 11971
–
11984, S
e
p. 2023, doi:
10.1007/s
12652
-
022
-
03748
-
6.
[
4]
H
.
T
a
he
r
doos
t
a
nd
M
. M
a
da
nc
hi
a
n,
“
A
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
a
nd
s
e
nt
im
e
nt
a
na
ly
s
is
:
a
r
e
vi
e
w
in
c
ompe
ti
ti
ve
r
e
s
e
a
r
c
h,”
C
om
put
e
r
s
,
vol
. 12, no. 2, F
e
b. 2023, doi:
10.3390/computer
s
12020037.
[
5]
N
.
J
e
ong
a
nd
J
.
L
e
e
,
“
A
n
a
s
pe
c
t
-
ba
s
e
d
r
e
vi
e
w
a
na
ly
s
i
s
us
in
g
C
ha
tG
P
T
f
or
th
e
e
xpl
or
a
ti
on
of
hot
e
l
s
e
r
vi
c
e
f
a
il
ur
e
s
,”
Sus
ta
in
abi
li
ty
, vol
. 16, no. 4, F
e
b. 2024, doi:
10.3390/s
u16041640.
[
6]
Y
.
C
.
H
ua
,
P
.
D
e
nny,
K
.
T
a
s
kova
,
a
nd
J
.
W
ic
ke
r
,
“
A
s
ys
te
m
a
ti
c
r
e
vi
e
w
of
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
(
A
B
S
A
)
:
doma
in
s
,
me
th
ods
, a
nd t
r
e
nds
,”
a
r
X
iv
e
-
pr
in
ts
, ar
X
iv
-
2311
, 2023.
[
7]
G
.
D
’
A
ni
e
ll
o,
M
.
G
a
e
ta
,
a
nd
I
.
L
a
R
oc
c
a
,
“
K
now
M
I
S
-
A
B
S
A
:
a
n
ove
r
vi
e
w
a
nd
a
r
e
f
e
r
e
nc
e
mode
l
f
o
r
a
ppl
ic
a
ti
ons
of
s
e
nt
i
me
nt
a
na
ly
s
is
a
nd
a
s
p
e
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,”
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
R
e
v
ie
w
,
vol
.
55,
no.
7,
pp.
5543
–
5574,
O
c
t.
2022,
doi
:
10.1007/s
10462
-
021
-
10134
-
9.
[
8]
R
.
K
. Y
a
da
v,
L
.
J
ia
o,
M
.
G
oodwin,
a
nd
O
.
-
C
.
G
r
a
nmo,
“
P
os
it
io
nl
e
s
s
a
s
pe
c
t
ba
s
e
d s
e
nt
im
e
nt
a
na
ly
s
i
s
us
in
g
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m,”
K
now
le
dge
-
B
as
e
d Sy
s
te
m
s
, vol
. 226, Aug. 2021, do
i:
10.1016/j
.
knos
ys
.2021.107136.
[
9]
S
.
K
a
lb
hor
a
nd
D
.
G
oya
l,
“
S
ur
ve
y
on
A
B
S
A
ba
s
e
d
on
ma
c
h
in
e
le
a
r
ni
ng,
de
e
p
le
a
r
ni
ng
a
nd
tr
a
ns
f
e
r
le
a
r
ni
ng
a
ppr
oa
c
h,”
A
I
P
C
onf
e
r
e
nc
e
P
r
oc
e
e
di
ng
s
, 2023, doi:
10.1063/5.0154549
.
[
10]
B
.
H
a
r
jo
a
nd
R
.
A
bdul
la
h,
“
A
tt
e
nt
io
n
-
ba
s
e
d
s
e
nt
e
n
c
e
e
xt
r
a
c
ti
o
n
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
i
s
w
it
h
im
pl
ic
it
a
s
pe
c
t
c
a
s
e
s
in
hot
e
l
r
e
vi
e
w
us
in
g
ma
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm,
s
e
ma
nt
ic
s
im
il
a
r
it
y,
a
nd
B
E
R
T
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
I
nt
e
ll
ig
e
nt
E
ngi
ne
e
r
in
g
and Sy
s
te
m
s
, vol
. 16, no. 3, pp. 189
–
200,
J
un. 2023, doi:
10.22266/i
ji
e
s
2023.0630.15.
[
11]
K
.
W
.
T
r
is
na
a
nd
H
.
J
.
J
ie
,
“
D
e
e
p
le
a
r
ni
ng
a
ppr
oa
c
h
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
c
la
s
s
if
ic
a
ti
on:
a
c
ompa
r
a
ti
ve
r
e
vi
e
w
,”
A
ppl
ie
d
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
, vol
. 36, no. 1, De
c
. 2022, doi:
10.1080/08839514.
2021.2014186.
[
12]
H
.
H
.
D
o,
P
.
W
.
C
.
P
r
a
s
a
d,
A
.
M
a
a
g,
a
nd
A
.
A
ls
a
doon,
“
D
e
e
p
le
a
r
ni
ng
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
:
a
c
ompa
r
a
ti
ve
r
e
vi
e
w
,”
E
x
pe
r
t
Sy
s
t
e
m
s
w
it
h A
ppl
ic
at
io
ns
, vol
. 118, pp. 272
–
2
99, M
a
r
. 2019, doi:
10.1016/j
.e
s
w
a
.2018.10.003.
[
13]
W
.
G
.
H
a
tc
he
r
a
nd
W
.
Y
u,
“
A
s
ur
ve
y
of
d
e
e
p
le
a
r
ni
ng:
pl
a
t
f
or
ms
,
a
ppl
ic
a
ti
ons
a
nd
e
m
e
r
gi
ng
r
e
s
e
a
r
c
h
tr
e
nds
,”
I
E
E
E
A
c
c
e
s
s
,
vol
. 6, pp. 24411
–
24432, 2018, doi:
10.1109/AC
C
E
S
S
.2018.28
30661.
[
14]
S
.
R
a
z
a
vi
,
“
D
e
e
p
le
a
r
ni
ng,
e
xpl
a
in
e
d:
F
unda
me
nt
a
ls
,
e
x
pl
a
in
a
bi
li
ty
,
a
nd
br
id
ge
a
bi
li
ty
to
pr
oc
e
s
s
-
ba
s
e
d
mod
e
ll
in
g,”
E
nv
ir
onm
e
nt
al
M
ode
ll
in
g & Sof
tw
ar
e
, vol
. 144, Oc
t.
2021, doi:
10.1016/j
.e
nvs
of
t.
2021.105159.
[
15]
J
.
S
un,
P
.
H
a
n,
Z
.
C
he
ng,
E
.
W
u,
a
nd
W
.
W
a
ng,
“
T
r
a
ns
f
or
me
r
ba
s
e
d
mul
ti
-
gr
a
in
e
d
a
tt
e
nt
io
n
ne
twor
k
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 8, pp. 211152
–
211163, 2020, doi:
10.1109/AC
C
E
S
S
.2020.3039470.
[
16]
B
.
Z
e
ng,
H
.
Y
a
ng,
R
.
X
u,
W
.
Z
hou,
a
nd
X
.
H
a
n,
“
L
C
F
:
a
lo
c
a
l
c
ont
e
xt
f
oc
us
me
c
ha
ni
s
m
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
c
la
s
s
if
ic
a
ti
on,”
A
ppl
ie
d Sc
ie
nc
e
s
, vol
. 9, no. 16, Aug. 201
9, doi:
10.3390/app916338
9.
[
17]
M
.
H
.
P
ha
n
a
nd
P
.
O
.
O
gunbona
,
“
M
od
e
ll
in
g
c
ont
e
xt
a
n
d
s
ynt
a
c
ti
c
a
l
f
e
a
tu
r
e
s
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
i
s
,”
in
P
r
oc
e
e
di
ngs
of
th
e
58t
h
A
nnual
M
e
e
ti
ng
of
th
e
A
s
s
oc
ia
ti
on
fo
r
C
om
put
at
io
nal
L
in
gui
s
ti
c
s
,
2020,
pp.
3211
–
3220,
doi
:
10.18653/v1/
2020.a
c
l
-
ma
in
.293.
[
18]
H
.
W
u,
Z
.
Z
ha
ng,
S
.
S
hi
,
Q
.
W
u,
a
nd
H
.
S
ong,
“
P
hr
a
s
e
de
pe
nde
nc
y
r
e
la
ti
ona
l
gr
a
ph
a
tt
e
nt
io
n
ne
twor
k
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
i
me
nt
a
na
ly
s
is
,”
K
now
le
dge
-
B
as
e
d S
y
s
te
m
s
, vol
. 236, J
a
n. 2022, doi:
10.1016/j
.knos
ys
.2021.107736.
[
19]
J
.
H
e
,
A
.
W
uma
ie
r
,
Z
.
K
a
d
e
e
r
,
W
.
S
un,
X
.
X
in
,
a
nd
L
.
Z
he
ng,
“
A
lo
c
a
l
a
nd
gl
oba
l
c
ont
e
xt
f
oc
us
mul
ti
li
ngua
l
le
a
r
ni
ng
mode
l
f
or
a
s
pe
c
t
-
ba
s
e
d s
e
nt
im
e
nt
a
n
a
ly
s
is
,
”
I
E
E
E
A
c
c
e
s
s
, vol
. 10, pp. 841
35
–
84146, 2022, doi:
10.1109/AC
C
E
S
S
.2022.3197218.
[
20]
B
.
Y
u
a
nd
S
.
Z
ha
ng,
“
A
nove
l
w
e
ig
ht
-
or
ie
nt
e
d
g
r
a
ph
c
onvolut
i
ona
l
ne
twor
k
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,”
T
he
J
our
na
l
of
Supe
r
c
om
put
in
g
, vol
. 79, no. 1, pp. 947
–
972, J
a
n. 2023, doi:
10
.1007/s
11227
-
022
-
04689
-
9.
[
21]
Y
.
M
a
,
R
.
S
ong,
X
.
G
u,
Q
.
S
he
n,
a
nd
H
.
X
u,
“
M
ul
ti
pl
e
gr
a
ph
c
onvolut
io
na
l
ne
twor
ks
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,”
A
ppl
ie
d I
nt
e
ll
ig
e
nc
e
, vol
. 53, no. 10, pp. 12985
–
12998, M
a
y 2023, doi:
10.1007/s
10489
-
022
-
04023
-
z.
[
22]
T
.
L
in
a
nd
I
.
J
oe
,
“
A
n
a
da
pt
iv
e
ma
s
ke
d
a
tt
e
nt
io
n
me
c
ha
ni
s
m
to
a
c
t
on
th
e
lo
c
a
l
te
xt
in
a
gl
ob
a
l
c
ont
e
xt
f
or
a
s
pe
c
t
-
ba
s
e
d
s
e
nt
i
me
nt
a
na
ly
s
is
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 11, pp. 43055
–
43066, 2023, doi:
1
0.1109/AC
C
E
S
S
.2023.3270927.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
14
,
No.
6
,
De
c
e
mber
20
24
:
6689
-
6698
6698
[
23]
Q
.
Z
ha
o,
F
.
Y
a
ng,
D
.
A
n,
a
nd
J
.
L
ia
n,
“
M
ode
li
ng
s
tr
uc
tu
r
e
d
d
e
pe
nde
nc
y
tr
e
e
w
it
h
gr
a
ph
c
onvolut
io
na
l
ne
twor
ks
f
or
a
s
p
e
c
t
-
le
ve
l
s
e
nt
im
e
nt
c
la
s
s
if
ic
a
ti
on,”
Se
ns
o
r
s
, vol
. 24, no. 2,
J
a
n. 2024, doi
:
10.3390/s
24020418.
[
24]
A
.
F
a
r
khod,
A
.
A
bdus
a
lo
mov,
F
.
M
a
khmudov,
a
nd
Y
.
I
.
C
ho,
“
L
D
A
-
ba
s
e
d
to
pi
c
mode
li
ng
s
e
nt
im
e
nt
a
na
ly
s
is
us
in
g
to
pi
c
/d
oc
ume
nt
/s
e
nt
e
nc
e
(
T
D
S
)
mode
l,
”
A
ppl
ie
d Sc
ie
nc
e
s
, vol
.
11, no. 23, Nov. 202
1, doi:
10.3390/app112311
091.
[
25]
K
.
S
.
A
r
un
a
nd
V
.
K
.
G
ovi
nda
n,
“
A
hybr
id
de
e
p
le
a
r
ni
ng
a
r
c
hi
te
c
tu
r
e
f
or
la
te
nt
to
pi
c
-
ba
s
e
d
im
a
ge
r
e
tr
ie
va
l,
”
D
at
a
Sc
ie
nc
e
and
E
ngi
ne
e
r
in
g
, vol
. 3, no. 2, pp. 166
–
195, J
un. 2018, doi:
10.1007/s
41019
-
018
-
0063
-
7.
[
26]
S
.
S
ye
d
a
nd
M
.
S
pr
ui
t,
“
F
ul
l
-
te
xt
or
a
bs
tr
a
c
t?
e
xa
mi
ni
ng
to
p
ic
c
ohe
r
e
nc
e
s
c
or
e
s
us
in
g
la
te
nt
D
ir
ic
hl
e
t
a
ll
oc
a
ti
on,”
in
2017
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
D
at
a
Sc
ie
nc
e
and
A
dv
anc
e
d
A
nal
y
ti
c
s
(
D
SA
A
)
,
O
c
t.
2017,
pp.
165
–
174,
doi
:
10.1109/DS
A
A
.2017.61.
[
27]
“
B
E
R
T
,”
H
uggi
ng F
ac
e
. ht
tp
s
:/
/h
uggi
ngf
a
c
e
.c
o/
doc
s
/t
r
a
ns
f
or
m
e
r
s
/e
n/
mode
l_
doc
/b
e
r
t
(
a
c
c
e
s
s
e
d A
ug. 30, 2024)
.
[
28]
P
r
a
c
to
,
“
B
ook
doc
to
r
a
ppoi
nt
me
nt
s
onl
in
e
,
o
r
de
r
me
di
c
in
e
,
di
a
gnos
ti
c
te
s
ts
,
c
ons
ul
t,
”
p
r
ac
to
.c
om
,
20
17.
ht
tp
s
:/
/ww
w
.pr
a
c
to
.c
om/
ba
nga
lo
r
e
/d
oc
to
r
s
(
a
c
c
e
s
s
e
d
A
ug. 30,
2024)
.
[
29]
“
I
ndi
a
’
s
la
r
ge
s
t
r
e
vi
e
w
pl
a
tf
or
m
,
”
m
out
hs
hut
.c
om
,
ht
tp
s
:/
/ww
w
.mout
hs
hut
.c
om/
c
a
te
gor
y/
ba
nga
lo
r
e
-
hos
pi
ta
ls
-
a
nd
-
c
li
ni
c
s
-
r
e
vi
e
w
s
-
925044773 (
a
c
c
e
s
s
e
d F
e
b. 13, 2024)
.
[
30]
M
.
P
ont
ik
i,
D
.
G
a
la
ni
s
,
J
.
P
a
vl
opoulos
,
H
.
P
a
pa
g
e
or
gi
ou,
I
.
A
ndr
out
s
opoulos
,
a
nd
S
.
M
a
na
ndh
a
r
,
“
S
e
mE
va
l
-
2014
T
a
s
k
4:
a
s
pe
c
t
ba
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
is
,”
in
P
r
oc
e
e
di
ngs
of
th
e
8t
h
I
nt
e
r
nat
io
nal
W
or
k
s
hop
on
Se
m
ant
ic
E
v
al
uat
io
n
(
Se
m
E
v
al
2014
)
,
2
014,
pp. 27
–
35, doi:
10.3115/v1/
S
14
-
2004.
[
31]
J
.
N
i,
J
.
L
i,
a
nd
J
.
M
c
A
ul
e
y,
“
J
us
ti
f
yi
ng
r
e
c
omm
e
nda
ti
ons
us
in
g
di
s
ta
nt
ly
-
la
be
le
d
r
e
vi
e
w
s
a
nd
f
in
e
-
gr
a
in
e
d
a
s
pe
c
ts
,
”
in
P
r
oc
e
e
di
ngs
of
th
e
2019
C
onf
e
r
e
nc
e
on
E
m
pi
r
ic
al
M
e
th
ods
in
N
at
ur
al
L
anguage
P
r
oc
e
s
s
in
g
and
th
e
9t
h
I
nt
e
r
nat
io
nal
J
oi
nt
C
onf
e
r
e
nc
e
on N
at
ur
al
L
anguage
P
r
oc
e
s
s
in
g (
E
M
N
L
P
-
I
J
C
N
L
P
)
, 2019, pp. 188
–
197, doi:
10.18653/v1/
D
19
-
1018.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
R
a
dhi
k
a
J
i
nendra
D
ha
n
a
l
i
s
a
n
acad
em
i
ci
a
n
fro
m
K
o
l
h
a
p
u
r,
Mah
ara
s
h
t
ra,
w
i
t
h
a
p
as
s
i
o
n
fo
r
co
m
p
u
t
er
s
ci
en
ce
an
d
i
n
f
o
rmat
i
o
n
s
y
s
t
em
s
.
Sh
e
i
s
cu
rren
t
l
y
affi
l
i
at
e
d
w
i
t
h
D
Y
Pat
i
l
Co
l
l
e
g
e
o
f
E
n
g
i
n
eer
i
n
g
a
n
d
T
ech
n
o
l
o
g
y
,
s
i
t
u
a
t
ed
i
n
K
as
a
b
a
Ba
w
ad
a,
K
o
l
h
a
p
u
r,
w
h
er
e
s
h
e
h
as
b
e
en
co
n
t
r
i
b
u
t
i
n
g
t
o
t
h
e
fi
e
l
d
o
f
ed
u
cat
i
o
n
s
i
n
ce
2
0
0
3
.
H
er
acad
emi
c
j
o
u
r
n
ey
l
e
d
h
er
t
o
p
u
r
s
u
e
a
mas
t
er's
i
n
e
n
g
i
n
eer
i
n
g
(M.
E
.
)
fro
m
S
h
i
v
aj
i
U
n
i
v
er
s
i
t
y
i
n
2
0
1
2
,
s
p
eci
a
l
i
z
i
n
g
i
n
h
er
area
o
f
i
n
t
eres
t
.
H
er
co
mm
i
t
me
n
t
t
o
acad
em
i
c
ex
ce
l
l
e
n
ce
an
d
a
d
eep
u
n
d
er
s
t
an
d
i
n
g
o
f
co
mp
u
t
er
s
ci
e
n
ce
an
d
i
n
fo
rma
t
i
o
n
s
y
s
t
em
s
l
ai
d
t
h
e
fo
u
n
d
at
i
o
n
fo
r
h
er
s
u
cces
s
fu
l
career.
W
i
t
h
a
ri
ch
acad
emi
c
b
ack
g
ro
u
n
d
,
a
w
eal
t
h
o
f
ex
p
er
i
en
c
e,
an
d
a
co
mmi
t
me
n
t
t
o
ad
v
an
c
i
n
g
t
h
e
fi
e
l
d
o
f
Co
m
p
u
t
er
Sci
e
n
ce,
Mi
s
s
Ra
d
h
i
k
a
D
h
an
a
l
s
t
an
d
s
as
a
n
o
t
ew
o
rt
h
y
fi
g
u
re
i
n
t
h
e
acad
em
i
c
l
an
d
s
ca
p
e
o
f
D
Y
Pat
i
l
C
o
l
l
eg
e
o
f
E
n
g
i
n
eeri
n
g
a
n
d
T
ech
n
o
l
o
g
y
,
K
o
l
h
a
p
u
r.
H
er
j
o
u
r
n
ey
co
n
t
i
n
u
e
s
t
o
i
n
s
p
i
re
b
o
t
h
s
t
u
d
e
n
t
s
an
d
fel
l
o
w
ed
u
cat
o
rs
al
i
k
e.
Sh
e
can
b
e
c
o
n
t
act
e
d
at
ema
i
l
:
d
h
a
n
al
l
rad
d
h
e
k
a@
g
mai
l
.
co
m
.
V
i
j
a
y
R
a
m
Gh
o
rpa
de
i
s
a
n
acad
em
i
c
res
earc
h
er
fr
o
m
Sh
i
v
aj
i
U
n
i
v
ers
i
t
y
.
H
e
i
s
t
h
e
Pr
o
fes
s
o
r
/
Pri
n
ci
p
al
o
f
B
h
art
i
V
i
d
y
ap
ee
t
h
’
s
C
o
l
l
eg
e
o
f
E
n
g
i
n
eer
i
n
g
,
K
o
l
h
ap
u
r.
He
h
a
s
co
n
t
ri
b
u
t
ed
t
o
re
s
earch
o
n
t
o
p
i
c
s
:
W
i
re
l
es
s
s
e
n
s
o
r
n
et
w
o
rk
s
an
d
Q
u
al
i
t
y
o
f
s
erv
i
ce.
He
h
as
a
H
-
i
n
d
e
x
o
f
7
,
co
-
au
t
h
o
re
d
2
4
p
u
b
l
i
ca
t
i
o
n
s
recei
v
i
n
g
1
4
6
c
i
t
a
t
i
o
n
s
.
H
e
can
b
e
co
n
t
a
ct
e
d
at
emai
l
:
v
i
j
ay
g
h
o
rp
a
d
e@
h
o
t
mai
l
.
co
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.