I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
2
,
A
p
r
il 2
0
2
6
,
p
p
.
1
8
1
4
~
1
8
2
4
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
8
1
4
-
1
8
2
4
1814
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Sema
ntic
-
sy
ntac
ti
c gra
ph net
wo
rk
f
o
r as
pect
-
ba
sed s
entimen
t
a
na
ly
sis
Rek
ha
B
du
rg
a
H
a
rish
1
,
Nee
la
m
bik
e
Sid
da
lin
g
a
ia
h
2
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
B
a
p
u
j
i
I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
D
a
v
a
n
a
g
e
r
e
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
G
M
U
n
i
v
e
r
s
i
t
y
,
D
a
v
a
n
a
g
e
r
e
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
3
0
,
2
0
2
5
R
ev
is
ed
Dec
2
6
,
2
0
2
5
Acc
ep
ted
J
an
2
5
,
2
0
2
6
As
p
e
c
t
-
b
a
se
d
se
n
ti
m
e
n
t
a
n
a
ly
si
s
(ABSA)
is
a
f
in
e
-
g
ra
i
n
e
d
se
n
ti
m
e
n
t
a
n
a
ly
sis
tas
k
th
a
t
i
d
e
n
ti
fies
se
n
ti
m
e
n
t
p
o
lariti
e
s
t
o
wa
rd
sp
e
c
ifi
c
a
sp
e
c
ts
with
in
a
se
n
ten
c
e
.
W
h
il
e
c
o
n
v
e
n
t
io
n
a
l
m
o
d
e
ls
h
a
v
e
a
c
h
iev
e
d
p
ro
g
re
ss
,
th
e
y
o
ften
n
e
g
lec
t
to
j
o
in
t
ly
c
o
n
sid
e
r
b
o
th
se
m
a
n
ti
c
c
o
n
tex
t
a
n
d
sy
n
tac
ti
c
stru
c
tu
re
,
li
m
it
in
g
p
e
rfo
rm
a
n
c
e
in
c
o
m
p
lex
li
n
g
u
isti
c
sc
e
n
a
rio
s.
Ne
v
e
rth
e
les
s,
m
o
st
e
x
isti
n
g
g
r
a
p
h
c
o
n
v
o
lu
t
io
n
a
l
n
e
two
r
k
(G
CN)
-
b
a
se
d
a
p
p
ro
a
c
h
e
s
h
a
v
e
re
c
e
n
tl
y
f
o
c
u
se
d
o
n
e
it
h
e
r
se
m
a
n
ti
c
o
r
sy
n
tac
ti
c
in
fo
rm
a
ti
o
n
i
n
d
i
v
i
d
u
a
ll
y
,
lea
d
i
n
g
to
su
b
o
p
ti
m
a
l
se
n
ti
m
e
n
t
c
l
a
ss
ifi
c
a
ti
o
n
a
c
c
u
ra
c
y
.
He
n
c
e
,
th
is
wo
rk
a
ims
to
d
e
si
g
n
a
n
e
ffe
c
ti
v
e
ABSA
m
o
d
e
l
th
a
t
sim
u
lt
a
n
e
o
u
sl
y
c
a
p
t
u
re
s
b
o
th
se
m
a
n
ti
c
re
latio
n
s
h
ip
s
a
n
d
sy
n
tac
ti
c
d
e
p
e
n
d
e
n
c
ies
fo
r
e
n
h
a
n
c
e
d
a
sp
e
c
t
-
lev
e
l
se
n
ti
m
e
n
t
a
n
a
ly
sis.
F
o
r
so
lv
i
n
g
issu
e
s
o
f
G
CN
-
b
a
se
d
a
p
p
ro
a
c
h
e
s,
th
is
wo
rk
p
ro
p
o
se
d
a
m
o
d
e
l
c
a
ll
e
d
se
n
ti
m
e
n
t
se
m
a
n
ti
c
sy
n
tac
ti
c
n
e
t
wo
rk
(
S
e
n
tS
e
m
S
y
n
Ne
t),
wh
ich
c
o
n
stru
c
ts
a
u
n
ifi
e
d
g
ra
p
h
b
y
in
teg
ra
ti
n
g
se
m
a
n
ti
c
a
n
d
sy
n
tac
ti
c
fe
a
tu
re
s
a
n
d
a
p
p
li
e
s
g
ra
p
h
n
e
u
ra
l
n
e
tw
o
rk
s
t
o
lea
rn
rich
,
a
sp
e
c
t
-
sp
e
c
ifi
c
re
p
re
se
n
tatio
n
s.
Th
e
m
o
d
e
l
wa
s
e
v
a
l
u
a
ted
o
n
t
h
e
S
e
m
Ev
a
l2
0
1
4
re
sta
u
ra
n
t
a
n
d
lap
t
o
p
d
a
tas
e
ts.
It
a
c
h
iev
e
d
8
8
.
2
5
%
a
c
c
u
ra
c
y
a
n
d
8
2
.
9
5
%
m
a
c
ro
-
F
-
sc
o
re
fo
r
re
sta
u
ra
n
t,
a
n
d
8
4
.
5
2
%
a
c
c
u
ra
c
y
a
n
d
8
0
.
2
6
%
m
a
c
ro
-
F
-
sc
o
re
fo
r
lap
t
o
p
.
T
h
e
m
o
d
e
l’s
u
n
iq
u
e
in
teg
ra
ti
o
n
o
f
b
o
th
se
m
a
n
ti
c
a
n
d
sy
n
tac
ti
c
imp
o
rtan
c
e
t
h
ro
u
g
h
a
u
n
ifi
e
d
g
ra
p
h
str
u
c
tu
re
im
p
ro
v
e
d
se
n
ti
m
e
n
t
d
e
tec
ti
o
n
a
c
c
u
ra
c
y
.
K
ey
w
o
r
d
s
:
Asp
e
ct
-
b
ase
d
s
e
n
ti
m
e
n
t
a
n
al
y
s
is
Dee
p
lear
n
in
g
Gr
ap
h
n
e
u
r
al
n
etwo
r
k
Sem
an
tic
-
s
y
n
tactic
f
u
s
io
n
Sen
tim
en
t c
lass
if
icatio
n
Sen
tSem
Sy
n
Net
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
R
ek
h
a
B
d
u
r
g
a
Har
is
h
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atio
n
Scien
ce
an
d
E
n
g
in
ee
r
i
n
g
,
B
ap
u
ji
I
n
s
titu
te
o
f
E
n
g
in
ee
r
in
g
an
d
T
ec
h
n
o
lo
g
y
Dav
an
g
er
e,
I
n
d
ia
E
m
ail:
r
ek
h
ab
h
1
2
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Sen
tim
en
t
an
aly
s
is
,
also
k
n
o
wn
as
o
p
in
io
n
m
in
in
g
,
is
a
s
u
b
f
ield
o
f
n
atu
r
al
la
n
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
NL
P)
th
at
f
o
cu
s
es
o
n
i
d
en
tif
y
in
g
a
n
d
ca
te
g
o
r
izin
g
o
p
in
io
n
s
ex
p
r
ess
ed
in
te
x
t.
I
t
s
ee
k
s
to
d
eter
m
in
e
wh
eth
er
th
e
ex
p
r
ess
ed
s
en
tim
en
t
is
p
o
s
itiv
e,
n
eg
ativ
e,
o
r
n
eu
tr
al.
Sen
tim
en
t
an
aly
s
is
h
as
b
ec
o
m
e
a
cr
u
cial
co
m
p
o
n
en
t
in
v
ar
i
o
u
s
r
ea
l
-
wo
r
ld
ap
p
licati
o
n
s
s
u
ch
as
p
r
o
d
u
ct
r
e
v
iews,
s
o
cial
m
ed
ia
m
o
n
ito
r
i
n
g
,
an
d
cu
s
to
m
er
f
ee
d
b
ac
k
s
y
s
tem
s
[
1
]
−
[
4
]
.
W
h
ile
tr
ad
i
tio
n
al
s
en
tim
en
t
an
aly
s
is
ev
alu
ates
th
e
o
v
er
all
s
en
tim
en
t
o
f
a
s
en
te
n
ce
o
r
d
o
cu
m
e
n
t,
it
o
f
ten
o
v
er
lo
o
k
s
th
e
f
ac
t
th
at
s
en
tim
en
ts
ca
n
v
ar
y
ac
r
o
s
s
d
if
f
er
en
t
asp
ec
ts
o
f
a
s
in
g
le
en
tity
.
T
h
is
lead
s
to
th
e
m
o
r
e
f
in
e
-
g
r
ain
e
d
task
o
f
asp
ec
t
-
b
ased
s
en
tim
en
t
an
aly
s
is
(
AB
SA)
.
A
B
SA
aim
s
to
d
etec
t
th
e
s
en
tim
en
t
ex
p
r
ess
ed
to
war
d
s
p
ec
if
ic
asp
ec
ts
o
r
attr
ib
u
tes
o
f
an
e
n
tity
m
e
n
tio
n
ed
in
a
s
en
ten
ce
[
5
]
−
[
7
]
.
Fo
r
ex
am
p
le,
in
th
e
r
ev
iew,
“T
h
e
s
cr
ee
n
is
b
r
ig
h
t,
b
u
t
th
e
b
atter
y
life
is
d
is
ap
p
o
in
tin
g
,
”
th
er
e
ar
e
two
asp
ec
ts
:
s
cr
ee
n
an
d
b
atter
y
life
.
A
g
en
er
al
s
en
tim
en
t
class
if
icatio
n
m
ig
h
t
in
co
r
r
ec
tly
la
b
el
th
e
s
e
n
ten
ce
as
n
e
u
tr
al
d
u
e
to
th
e
co
n
f
lictin
g
s
en
tim
en
ts
.
AB
SA,
h
o
wev
er
,
co
r
r
ec
tly
id
en
tifie
s
th
at
th
e
s
en
tim
en
t
to
war
d
s
s
cr
ee
n
is
p
o
s
itiv
e,
wh
ile
th
e
s
en
tim
en
t t
o
war
d
s
b
atter
y
life
is
n
e
g
ativ
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
ema
n
tic
-
s
yn
ta
ctic
g
r
a
p
h
n
et
w
o
r
k
fo
r
a
s
p
ec
t
-
b
a
s
ed
s
en
timen
t a
n
a
lysi
s
(
R
ek
h
a
B
d
u
r
g
a
Ha
r
is
h
)
1815
Mo
r
eo
v
er
,
a
cr
itical
ch
allen
g
e
in
AB
SA
lies
in
th
e
n
ee
d
to
u
n
d
er
s
tan
d
b
o
th
s
em
an
tic
an
d
s
y
n
tactic
r
elatio
n
s
h
ip
s
in
a
s
en
ten
ce
[
8
]
,
[
9
]
.
Sem
an
tic
in
f
o
r
m
atio
n
h
elp
s
ca
p
tu
r
e
th
e
m
ea
n
i
n
g
o
f
wo
r
d
s
an
d
th
eir
co
n
tex
tu
al
u
s
ag
e,
wh
ile
s
y
n
tac
tic
in
f
o
r
m
atio
n
f
o
cu
s
es
o
n
g
r
a
m
m
atica
l
s
tr
u
ctu
r
e
an
d
t
h
e
r
el
atio
n
s
h
ip
s
b
etwe
en
d
if
f
er
en
t
p
ar
ts
o
f
a
s
en
te
n
ce
.
Fo
r
in
s
tan
ce
,
in
th
e
s
en
ten
c
e,
“Alth
o
u
g
h
th
e
p
e
r
f
o
r
m
an
c
e
o
f
th
e
la
p
to
p
is
im
p
r
ess
iv
e,
th
e
d
esig
n
f
ee
ls
o
u
td
ated
,
”
th
e
m
o
d
el
m
u
s
t
s
em
an
tically
u
n
d
er
s
tan
d
th
at
“p
er
f
o
r
m
an
ce
”
an
d
“d
esig
n
”
ar
e
asp
ec
ts
,
an
d
s
y
n
tactica
lly
d
eter
m
in
e
wh
ic
h
s
en
tim
en
t
wo
r
d
s
ap
p
l
y
to
wh
ich
asp
ec
ts
.
Fo
r
ex
am
p
le,
wh
en
co
n
s
id
er
in
g
a
s
en
ten
ce
,
“T
h
e
ca
m
er
a
q
u
ality
o
f
th
is
p
h
o
n
e
is
ex
ce
llen
t
,
”
it
i
s
r
esp
ec
tiv
e
asp
ec
t
is
ca
m
er
a
q
u
ality
,
a
n
d
th
e
m
o
d
el
will
p
r
ed
ict
as
p
o
s
itiv
e.
Si
m
ilar
ly
,
f
o
r
an
o
t
h
er
s
en
ten
ce
,
“T
h
e
b
atter
y
d
r
ain
s
to
o
q
u
ick
ly
,
”
it
i
s
r
esp
ec
tiv
e
asp
ec
t
is
b
atter
y
an
d
th
e
AB
SA
m
o
d
el
will
p
r
ed
ict
as
n
eg
ativ
e.
I
n
m
an
y
ca
s
es,
a
s
in
g
le
s
en
ten
ce
co
n
tain
s
m
u
ltip
le
asp
ec
ts
,
ea
ch
with
its
o
wn
s
en
tim
en
t.
T
h
is
m
ak
es
o
v
er
all
s
en
tim
en
t
class
if
icatio
n
in
s
u
f
f
icien
t
f
o
r
t
ask
s
lik
e
o
p
in
io
n
s
u
m
m
ar
izatio
n
o
r
p
r
o
d
u
ct
im
p
r
o
v
em
en
t.
F
o
r
ex
am
p
le,
in
th
e
s
en
ten
ce
,
“T
h
e
d
is
p
lay
is
s
tu
n
n
in
g
,
b
u
t
th
e
s
p
ea
k
er
s
a
r
e
m
e
d
io
cr
e,
”
tr
ea
tin
g
th
is
as
a
s
in
g
le
s
en
tim
en
t
wo
u
ld
m
is
lead
d
ec
is
io
n
-
m
ak
er
s
.
I
n
s
t
ea
d
,
AB
SA
ca
p
tu
r
es
b
o
t
h
as
p
ec
ts
as
d
is
p
lay
b
ein
g
p
o
s
itiv
e
an
d
s
p
ea
k
er
s
as
n
eg
ativ
e.
T
o
ad
d
r
ess
s
u
ch
co
m
p
lex
ity
,
it
is
cr
u
cial
to
is
o
late
an
d
an
aly
ze
th
e
p
o
lar
ity
o
f
ea
ch
asp
ec
t
in
d
ep
en
d
en
tly
.
AB
SA
f
r
am
e
wo
r
k
s
d
o
th
is
b
y
lin
k
in
g
o
p
i
n
io
n
wo
r
d
s
(
e.
g
.
,
“stu
n
n
in
g
”
a
n
d
“m
e
d
io
cr
e”
)
with
th
eir
r
esp
ec
tiv
e
asp
ec
ts
b
ased
o
n
b
o
th
s
y
n
tactic
d
e
p
en
d
e
n
cies
an
d
s
em
an
tic
s
im
ilar
ities
.
T
h
is
ap
p
r
o
ac
h
en
s
u
r
es th
at
th
e
s
en
tim
en
t a
s
s
ig
n
ed
to
ea
ch
asp
ec
t is p
r
ec
is
e,
en
ab
lin
g
m
o
r
e
e
f
f
ec
tiv
e
s
en
ti
m
en
t a
n
aly
s
is
.
Als
o
,
in
r
ec
en
t
y
ea
r
s
,
d
ee
p
lear
n
in
g
(
DL
)
ap
p
r
o
ac
h
es
h
av
e
b
ec
o
m
e
th
e
d
o
m
in
an
t
p
ar
ad
ig
m
f
o
r
AB
SA
[
1
0
]
.
Mo
d
els
s
u
ch
as
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
[
1
1
]
,
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
[
1
2
]
,
an
d
tr
an
s
f
o
r
m
er
s
lik
e
b
id
ir
ec
tio
n
al
-
en
c
o
d
er
-
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
e
r
s
(
B
E
R
T
)
[
1
3
]
h
av
e
d
em
o
n
s
tr
ated
s
tr
o
n
g
p
er
f
o
r
m
a
n
ce
d
u
e
to
th
eir
a
b
ilit
y
to
c
ap
tu
r
e
r
ich
s
em
an
tic
co
n
tex
t
a
n
d
lear
n
h
ier
ar
c
h
ical
f
ea
tu
r
es.
T
h
ese
m
o
d
els
h
av
e
r
ep
lace
d
m
an
u
al
f
ea
tu
r
e
en
g
in
e
er
in
g
an
d
r
u
le
-
b
ased
s
y
s
tem
s
b
y
lear
n
in
g
d
ir
ec
tly
f
r
o
m
d
ata.
H
o
wev
er
,
DL
m
o
d
els
o
f
ten
s
tr
u
g
g
le
with
s
tr
u
ctu
r
al
u
n
d
er
s
tan
d
i
n
g
,
p
a
r
ticu
lar
ly
wh
en
it
co
m
es
to
lo
n
g
-
r
a
n
g
e
d
ep
en
d
en
cies
an
d
co
m
p
lex
s
en
te
n
ce
s
tr
u
ctu
r
es.
T
h
ey
m
a
y
ig
n
o
r
e
s
y
n
tactic
im
p
o
r
tan
ce
,
lead
in
g
to
m
is
class
if
icatio
n
o
f
s
en
tim
en
ts
as
s
o
ciate
d
with
s
p
ec
if
ic
as
p
ec
ts
.
T
o
ad
d
r
ess
th
is
,
g
r
ap
h
-
b
ased
DL
m
o
d
els
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
,
lev
er
a
g
in
g
g
r
a
p
h
co
n
v
o
lu
tio
n
al
n
e
two
r
k
s
(
GC
Ns)
[
1
4
]
,
[
1
5
]
to
en
co
d
e
s
y
n
tactic
d
ep
en
d
e
n
cies
u
s
in
g
d
ep
en
d
e
n
cy
p
ar
s
in
g
tr
ee
s
.
T
h
ese
m
o
d
els
allo
w
to
k
en
s
to
in
ter
ac
t
b
ased
o
n
th
eir
g
r
am
m
atica
l
r
elatio
n
s
h
ip
s
,
r
at
h
er
th
an
lin
ea
r
o
r
d
e
r
alo
n
e,
s
ig
n
if
ican
tly
im
p
r
o
v
in
g
asp
ec
t
-
s
e
n
tim
en
t
alig
n
m
e
n
t.
Desp
ite
th
eir
s
u
cc
ess
,
m
an
y
ex
is
tin
g
g
r
ap
h
-
b
ased
ap
p
r
o
ac
h
es
ten
d
to
f
o
c
u
s
ex
clu
s
iv
ely
o
n
eith
er
s
em
an
tic
o
r
s
y
n
tactic
in
f
o
r
m
atio
n
,
r
esu
ltin
g
in
s
u
b
o
p
tim
al
p
e
r
f
o
r
m
a
n
ce
wh
en
th
e
two
f
o
r
m
s
o
f
i
n
f
o
r
m
atio
n
ar
e
n
o
t
jo
in
tly
co
n
s
id
er
e
d
.
T
o
o
v
er
c
o
m
e
th
is
lim
itatio
n
,
th
is
p
r
o
p
o
s
es
a
u
n
if
ied
ar
c
h
itectu
r
e
ca
lled
s
en
tim
en
t
s
em
an
tic
s
y
n
tactic
n
etwo
r
k
(
Sen
tSem
Sy
n
Net)
.
T
h
is
m
o
d
el
s
ea
m
less
ly
in
teg
r
at
es
d
ee
p
s
em
an
tic
r
ep
r
esen
tatio
n
s
f
r
o
m
B
E
R
T
with
tem
p
o
r
al
m
o
d
elin
g
th
r
o
u
g
h
b
i
d
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
‑
ter
m
m
em
o
r
y
(
B
iLST
M
)
an
d
s
y
n
tactic
s
tr
u
ctu
r
e
u
s
in
g
a
two
-
lay
er
GC
N.
T
h
e
u
s
e
o
f
B
E
R
T
ca
p
tu
r
es
wo
r
d
m
ea
n
in
g
in
co
n
tex
t,
B
iLST
M
m
o
d
els
t
h
e
wo
r
d
-
o
r
d
e
r
an
d
s
eq
u
en
ce
d
y
n
am
ics,
a
n
d
GC
N
p
r
o
p
ag
ates
s
y
n
tactic
d
e
p
en
d
en
cies.
B
y
co
m
b
in
in
g
th
ese
co
m
p
o
n
en
ts
,
Sen
tSem
Sy
n
Net
ac
h
iev
es
a
h
o
lis
tic
u
n
d
er
s
tan
d
in
g
o
f
b
o
th
m
ea
n
in
g
a
n
d
s
tr
u
ct
u
r
e,
en
ab
lin
g
m
o
r
e
ac
c
u
r
ate
s
en
tim
en
t c
lass
if
icatio
n
f
o
r
ea
ch
asp
ec
t.
T
h
e
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
as f
o
llo
ws
.
T
h
is
wo
r
k
p
r
o
p
o
s
es
a
n
o
v
el
h
y
b
r
id
m
o
d
el,
Sen
tSem
Sy
n
N
et,
th
at
in
teg
r
ates
B
E
R
T
,
B
iL
STM
,
an
d
GC
N
to
jo
in
tly
m
o
d
el
s
e
m
an
tic
an
d
s
y
n
tactic
in
f
o
r
m
atio
n
.
T
h
e
Sen
tSem
Sy
n
Net
m
o
d
el
id
en
tifie
s
asp
ec
t
-
s
p
ec
if
ic
s
en
tim
en
t
b
y
ca
p
tu
r
in
g
d
ee
p
co
n
tex
t
an
d
g
r
am
m
atica
l
s
tr
u
ctu
r
e.
T
h
e
Sen
tSem
Sy
n
Net
ar
ch
itectu
r
e
o
u
tp
er
f
o
r
m
s
e
x
is
tin
g
AB
SA
an
d
GC
N
ap
p
r
o
ac
h
es
f
o
r
AB
SA
d
atasets
in
ter
m
s
o
f
ac
c
u
r
ac
y
an
d
m
ac
r
o
-
F
-
s
co
r
e.
T
h
e
m
an
u
s
cr
ip
t
is
o
r
g
a
n
ized
in
th
e
f
o
llo
w
in
g
m
a
n
n
er
,
s
ec
tio
n
2
d
is
cu
s
s
es
ex
is
tin
g
AB
SA
ap
p
r
o
ac
h
es,
s
ec
tio
n
3
p
r
esen
t
s
th
e
Sen
tSem
Sy
n
Net
f
o
r
AB
SA
,
s
ec
tio
n
4
d
is
cu
s
s
es
th
e
r
esu
lts
ac
h
iev
ed
b
y
Sen
tSem
Sy
n
Net
an
d
it
co
m
p
a
r
es
with
ex
is
tin
g
ap
p
r
o
ac
h
es p
r
esen
ted
in
liter
atu
r
e
s
u
r
v
ey
,
a
n
d
f
in
ally
s
ec
tio
n
5
p
r
esen
ts
th
e
co
n
clu
s
io
n
an
d
f
u
tu
r
e
wo
r
k
.
2.
L
I
T
E
R
AT
U
RE
SU
RVE
Y
T
h
is
s
ec
tio
n
d
is
cu
s
s
es
th
e
e
x
is
tin
g
AB
S
A
ap
p
r
o
ac
h
es
p
r
esen
ted
in
r
ec
en
t
y
ea
r
s
f
o
r
s
en
tim
en
t
an
aly
s
is
.
Hu
an
g
et
a
l.
[
1
6
]
p
r
esen
ted
an
asp
ec
t
-
lev
el
s
en
tim
en
t
-
an
aly
s
is
ap
p
r
o
ac
h
,
ca
lled
co
n
tex
t
-
p
o
s
itio
n
-
awa
r
e
s
en
tim
en
t
an
aly
s
is
(
C
P
A
-
SA)
wh
er
e
m
ain
f
o
cu
s
was
to
ac
h
iev
e
th
e
asp
ec
t
-
s
p
ec
if
ic
co
n
tex
tu
al
-
lo
ca
tio
n
in
f
o
r
m
atio
n
.
I
n
th
eir
wo
r
k
,
tw
o
asy
m
m
etr
ical
co
n
tex
t
-
p
o
s
itio
n
weig
h
t
-
f
u
n
ctio
n
s
wer
e
d
es
ig
n
ed
,
w
h
er
ein
t
h
e
f
ir
s
t
weig
h
t
f
u
n
ctio
n
ad
ju
s
ted
co
n
tex
tu
al
wo
r
d
weig
h
ts
ac
c
o
r
d
in
g
to
asp
ec
t
wo
r
d
p
o
s
itio
n
s
in
s
en
ten
ce
s
an
d
th
e
s
ec
o
n
d
weig
h
t
f
u
n
ctio
n
i
m
p
r
o
v
e
d
th
e
s
en
tim
en
t
p
o
lar
i
ty
ju
d
g
em
e
n
t.
Fu
r
th
er
,
in
th
e
ir
wo
r
k
,
th
e
y
u
s
ed
m
u
ltip
le
an
d
s
in
g
le
s
en
ten
ce
-
lev
el
b
id
ir
ec
tio
n
al
g
ated
r
ec
u
r
r
en
t
u
n
it
(
B
iGR
U)
f
o
r
ex
tr
a
ctin
g
th
e
im
p
ac
t
o
f
co
n
tex
t
-
r
elate
d
r
elatio
n
s
h
ip
o
f
ev
er
y
s
en
ten
ce
in
d
ataset
o
n
b
asis
o
f
asp
ec
t
-
s
en
tim
en
t p
o
lar
i
ty
.
Mo
r
e
o
v
er
,
th
e
y
p
r
esen
ted
a
lo
s
s
f
u
n
ctio
n
f
o
r
h
an
d
lin
g
class
im
b
alan
ce
i
s
s
u
es.
E
v
alu
atio
n
s
wer
e
c
o
n
d
u
cted
c
o
n
s
id
er
in
g
Sem
E
v
al
d
atasets
,
wh
er
e
co
n
s
id
er
ed
R
estau
r
at2
0
1
6
,
R
estau
r
an
t2
0
1
5
,
R
estau
r
an
t2
0
1
4
,
an
d
L
ap
to
p
2
0
1
4
.
Fin
d
in
g
s
s
h
o
w
th
at
th
e
C
PA
-
SA
ap
p
r
o
ac
h
ac
h
iev
e
d
8
9
.
0
2
%,
7
9
.
6
1
%,
7
5
.
1
8
%
,
an
d
8
2
.
6
4
%
ac
cu
r
ac
y
f
o
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
8
1
4
-
1
8
2
4
1816
R
estau
r
at2
0
1
6
,
R
estau
r
an
t2
0
1
5
,
R
estau
r
an
t2
0
1
4
,
an
d
L
a
p
to
p
2
0
1
4
d
atasets
r
esp
ec
tiv
e
ly
.
L
u
et
a
l.
[
1
7
]
p
r
esen
ted
an
ap
p
r
o
ac
h
ca
lled
h
eter
o
g
en
o
u
s
-
g
r
ap
h
n
e
u
r
al
-
n
et
wo
r
k
(
HGNN
)
,
wh
er
e
th
ey
co
n
s
id
er
ed
in
ter
ac
tiv
e
asp
ec
t
co
n
tex
ts
a
n
d
w
o
r
d
s
f
o
r
en
c
o
d
in
g
s
en
ten
ce
s
eq
u
en
c
e
d
ata
f
o
r
p
ar
a
m
eter
s
h
ar
i
n
g
.
T
h
e
m
ain
aim
o
f
HGNN
was
to
en
co
d
e
s
y
n
tax
d
ep
en
d
e
n
cy
-
tr
ee
,
en
co
d
e
f
ew
o
f
p
ar
t
-
of
-
s
p
ee
ch
(
Po
S)
tag
s
a
n
d
en
co
d
e
p
r
ev
i
o
u
s
s
en
tim
en
t
-
d
ictio
n
ar
y
f
o
r
s
en
ti
m
en
t
p
r
e
d
ictio
n
.
E
v
alu
atio
n
s
wer
e
co
n
d
u
cted
o
n
f
iv
e
Sem
E
v
al
d
atasets
,
wh
er
e
f
in
d
in
g
s
s
h
o
w
th
at
HGNN
ap
p
r
o
ac
h
ac
h
iev
ed
8
7
.
9
2
%
,
8
0
.
3
7
%,
7
4
.
0
8
%,
8
1
.
9
1
%
,
an
d
7
7
.
2
9
%
f
o
r
R
estau
r
at2
0
1
6
,
R
estau
r
an
t2
0
1
5
,
R
estau
r
an
t2
0
1
4
,
L
ap
t
o
p
2
0
1
4
,
an
d
T
witter
2
0
1
4
d
atasets
r
esp
ec
tiv
ely
.
L
in
an
d
J
o
e
[
1
8
]
p
r
esen
ted
a
n
ap
p
r
o
ac
h
o
n
b
asis
o
f
m
ec
h
an
is
m
o
f
g
lo
b
al
an
d
m
ask
e
d
atten
tio
n
,
ca
lled
lo
ca
l
-
g
lo
b
al
co
n
tex
t
-
f
o
cu
s
(
L
GC
F).
I
n
th
eir
wo
r
k
,
th
ey
u
s
ed
p
r
ev
io
u
s
p
r
esen
ted
ap
p
r
o
ac
h
es
ca
lled
co
n
tex
t
-
f
ea
tu
r
es
d
y
n
am
ic
-
wei
g
h
t
an
d
c
o
n
tex
t
-
f
ea
tu
r
e
d
y
n
a
m
ic
m
ask
[
1
9
]
,
f
o
r
ass
ig
n
in
g
tex
t
-
v
ec
to
r
weig
h
ts
o
n
b
asis
o
f
d
is
tan
ce
f
r
o
m
a
s
p
ec
t
-
ter
m
.
Fu
r
th
er
,
p
r
o
p
o
s
e
d
an
ap
p
r
o
ac
h
wh
ich
u
tili
ze
d
m
ask
ed
-
atten
tio
n
ap
p
r
o
ac
h
f
o
r
in
ter
ce
p
tin
g
lo
ca
l
em
b
ed
d
in
g
with
in
g
lo
b
al
em
b
ed
d
in
g
an
d
th
en
e
v
alu
ated
as
p
ec
t
-
ter
m
p
o
s
itio
n
an
d
f
in
ally
r
eo
r
d
er
ed
weig
h
ts
o
n
b
asis
o
f
asp
ec
t
-
p
o
s
itio
n
an
d
ass
ig
n
ed
a
g
lo
b
al
em
b
ed
d
in
g
o
n
b
asis
o
f
its
r
esp
ec
tiv
e
s
u
b
s
cr
ip
ts
,
m
ak
in
g
th
e
m
o
d
el
to
ex
tr
ac
t m
o
r
e
f
ea
t
u
r
es a
n
d
p
r
o
v
id
e
a
n
o
is
e
f
r
ee
a
p
p
r
o
ac
h
.
Usi
n
g
th
e
f
o
llo
win
g
ap
p
r
o
ac
h
,
th
e
L
GC
F
ap
p
r
o
ac
h
lea
r
n
t
b
o
th
lo
ca
l
an
d
g
lo
b
al
f
ea
tu
r
es,
p
r
o
v
i
d
in
g
b
etter
s
en
tim
en
t
an
aly
s
is
p
er
f
o
r
m
an
ce
.
E
v
alu
atio
n
s
wer
e
co
n
d
u
cte
d
o
n
e
ig
h
t
d
atasets
,
wh
ich
in
clu
d
e
d
R
estau
r
an
t2
0
1
6
,
R
estau
r
asn
t2
0
1
4
,
L
ap
t
o
p
2
0
1
4
,
T
witter
2
0
1
4
,
an
d
m
u
lti
-
asp
e
ct
m
u
lti
-
s
en
tim
en
t
(
MA
MS)
(
ca
m
er
a,
ca
r
,
p
h
o
n
e,
telev
is
io
n
,
an
d
t
-
s
h
ir
t
)
d
ata.
Fo
r
MA
MS
d
ata,
th
e
L
G
C
F
ac
h
iev
ed
9
7
.
2
4
%,
9
8
.
2
6
%,
9
7
.
5
9
,
9
1
.
6
1
%
,
an
d
9
3
.
8
6
%
ac
c
u
r
ac
y
f
o
r
ca
m
er
a,
ca
r
,
p
h
o
n
e,
telev
is
io
n
,
an
d
t
-
s
h
ir
t
d
ata
r
esp
ec
tiv
ely
.
Fo
r
Se
m
E
v
al
d
ata,
ac
h
iev
ed
9
1
.
8
7
%,
8
5
.
5
2
%,
8
1
.
2
9
%
,
a
n
d
7
5
.
6
9
%
f
o
r
R
estau
r
an
t2
0
1
6
,
R
estau
r
an
t2
0
1
4
,
L
a
p
to
p
2
0
1
4
,
T
witter
2
0
1
4
r
esp
ec
tiv
ely
.
Z
h
a
o
et
a
l
.
[
2
0
]
p
r
esen
ted
an
a
p
p
r
o
ac
h
c
alled
as
s
tr
u
ctu
r
e
-
d
ep
e
n
d
en
c
y
tr
ee
-
b
ased
g
r
ap
h
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
(
SD
T
GC
N)
ap
p
r
o
ac
h
,
wh
ic
h
e
x
p
lo
r
ed
s
tr
u
ctu
r
e
d
s
y
n
tactic
d
ep
e
n
d
en
c
y
-
g
r
a
p
h
co
n
s
tr
u
ctio
n
c
o
n
s
id
er
in
g
Po
S
t
ag
s
,
s
en
tim
en
t
k
n
o
wled
g
e,
p
o
s
itio
n
in
f
o
r
m
atio
n
,
a
n
d
d
ep
en
d
en
cies
d
is
tan
ce
f
o
r
ass
ig
n
in
g
r
an
d
o
m
ed
g
e
-
b
ased
weig
h
ts
am
o
n
g
n
o
d
es.
Usi
n
g
t
h
e
g
r
ap
h
co
n
s
tr
u
ctio
n
,
th
e
co
n
n
ec
tio
n
am
o
n
g
th
e
im
p
o
r
tan
t
wo
r
d
s
an
d
asp
ec
t
n
o
d
es
in
cr
ea
s
es,
p
r
o
v
id
in
g
b
ett
er
s
en
tim
en
t
an
aly
s
is
.
I
n
th
eir
wo
r
k
,
u
s
ed
n
o
d
e
d
ep
en
d
e
n
cy
d
is
tan
ce
an
d
Po
S
tag
s
f
o
r
d
is
co
v
er
in
g
c
o
n
n
ec
ti
o
n
am
o
n
g
im
p
o
r
tan
t
n
o
d
es
w
ith
o
u
t
co
n
s
id
er
in
g
d
ir
ec
t
n
o
d
e
d
ep
e
n
d
en
c
y
.
Fu
r
th
er
,
th
e
g
r
ap
h
n
o
d
es
wer
e
a
g
g
r
eg
ated
o
n
b
asis
o
f
th
eir
in
f
o
r
m
atio
n
f
o
r
attain
in
g
ac
cu
r
ate
asp
ec
t
r
ep
r
esen
tatio
n
.
E
v
alu
atio
n
s
wer
e
co
n
d
u
ct
ed
o
n
f
iv
e
Sem
E
v
al
d
ataset
s
,
wh
ich
in
clu
d
ed
T
witter
2
0
1
4
,
R
estu
ar
an
t2
0
1
4
,
L
ap
to
p
2
0
1
4
,
R
estau
r
an
t2
0
1
5
,
an
d
R
estau
r
an
t2
0
1
6
,
w
h
er
e
ac
h
iev
ed
7
6
.
2
5
%,
8
3
.
8
2
%,
7
8
.
6
4
%,
8
3
.
2
1
%
,
a
n
d
9
1
.
5
3
% a
cc
u
r
ac
y
r
esp
ec
tiv
ely
.
Gu
et
a
l.
[
2
1
]
,
f
o
r
s
o
lv
in
g
is
s
u
es
o
f
GC
N,
wh
ich
f
ail
to
c
o
n
s
id
er
s
p
ec
if
ic
asp
ec
ts
in
s
en
ten
ce
f
o
r
s
en
tim
en
t
an
aly
s
is
,
p
r
esen
ted
a
m
o
d
el
ca
lled
s
y
n
tax
-
awa
r
e
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
SAGC
N)
.
T
h
e
SAGC
N
m
o
d
e
f
ir
s
t
u
tili
ze
d
asp
ec
t
-
s
p
ec
if
ic
f
ea
tu
r
es
in
c
o
n
tex
tu
al
i
n
f
o
r
m
atio
n
,
an
d
f
u
r
th
er
i
n
co
r
p
o
r
ated
ex
ter
n
al
s
en
tim
en
t
-
k
n
o
wled
g
e
f
o
r
im
p
r
o
v
in
g
GC
N
m
o
d
el
ca
p
ab
ilit
y
f
o
r
u
n
d
er
s
tan
d
in
g
s
en
tim
en
t in
f
o
r
m
atio
n
.
Fu
r
th
er
,
in
SAGC
N,
a
p
o
in
t
-
wis
e
co
n
v
o
lu
tio
n
-
tr
an
s
f
o
r
m
e
r
(
PC
T
)
an
d
m
u
lti
-
h
ea
d
s
elf
-
atten
tio
n
(
MH
SA)
ap
p
r
o
ac
h
wer
e
u
tili
ze
d
f
o
r
ca
p
tu
r
in
g
s
em
an
tic
in
f
o
r
m
atio
n
o
f
s
en
ten
ce
s
.
I
n
SAGC
N,
b
o
th
th
e
s
y
n
tactic
an
d
s
em
an
tic
s
en
ten
ce
in
f
o
r
m
atio
n
wer
e
co
n
s
id
er
ed
.
E
v
alu
atio
n
s
wer
e
co
n
d
u
cted
o
n
Sem
E
v
al2
0
1
4
d
ataset,
wh
er
e
ac
h
iev
ed
7
7
.
9
7
%,
9
7
.
5
3
%
,
an
d
8
3
.
0
6
%
ac
cu
r
ac
y
f
o
r
T
witter
2
0
1
4
,
R
estau
r
an
t2
0
1
4
,
an
d
L
ap
to
p
2
0
1
4
d
atase
t
r
esp
ec
tiv
ely
.
So
n
g
et
a
l.
[
2
2
]
p
r
esen
ted
an
a
p
p
r
o
ac
h
ca
lled
k
n
o
wled
g
e
-
g
u
id
ed
h
eter
o
g
en
o
u
s
g
r
a
p
h
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
(
KHGCN)
,
f
o
r
o
v
e
r
co
m
in
g
is
s
u
es
o
f
B
E
R
T
.
T
h
e
KHGCN
m
er
g
e
d
s
u
b
-
w
o
r
d
v
ec
to
r
s
u
s
in
g
d
y
n
am
ic
weig
h
ts
ap
p
r
o
ac
h
,
wh
ich
is
co
n
s
id
er
ed
in
B
E
R
T
em
b
ed
d
in
g
-
lay
er
.
Fu
r
th
er
,
h
ete
r
o
g
e
n
o
u
s
g
r
ap
h
s
we
r
e
b
u
ilt
f
o
r
f
u
s
in
g
v
ar
io
u
s
f
ea
tu
r
e
r
elatio
n
s
h
ip
am
o
n
g
wo
r
d
s
an
d
GC
N
was
u
tili
ze
d
f
o
r
i
d
en
tify
in
g
co
n
tex
t
-
s
p
ec
if
ic
s
y
n
tactic
f
ea
t
u
r
es.
Als
o
,
b
y
k
n
o
wled
g
e
-
g
r
a
p
h
em
b
ed
d
in
g
,
th
e
KHGCN
ap
p
r
o
ac
h
lear
n
t
m
o
r
e
f
ea
tu
r
es
f
r
o
m
d
if
f
er
e
n
t
s
o
u
r
c
es.
Usi
n
g
th
e
f
o
llo
win
g
k
n
o
wled
g
e
in
f
o
r
m
atio
n
,
th
e
s
y
n
t
ac
tic,
s
em
an
tic
an
d
k
n
o
wled
g
e
ex
tr
ac
te
d
f
ea
tu
r
es
wer
e
ag
g
r
eg
ated
p
r
o
v
id
in
g
a
f
ea
tu
r
e
-
f
u
s
io
n
ap
p
r
o
ac
h
f
o
r
s
en
tim
en
t
an
aly
s
is
.
T
h
e
KHGCN
wa
s
ev
alu
ated
u
s
in
g
R
estau
r
an
t2
0
1
6
,
R
estau
r
an
t2
0
1
5
,
an
d
L
ap
t
o
p
2
0
1
4
d
atas
et,
wh
er
e
ac
h
iev
ed
9
1
.
0
7
%,
8
5
.
4
2
%
,
an
d
8
0
.
8
7
%
ac
cu
r
ac
y
r
esp
ec
tiv
el
y
f
o
r
AB
SA.
C
h
en
et
a
l.
[
2
3
]
,
f
o
r
s
o
lv
in
g
is
s
u
es
o
f
u
n
d
er
s
tan
d
in
g
o
f
d
e
p
en
d
en
c
y
r
elatio
n
s
b
etwe
en
c
o
n
tex
tu
al
wo
r
d
s
,
th
is
wo
r
k
p
r
esen
ted
a
m
o
d
el
ca
lled
s
y
n
tactic
s
em
an
t
ic
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
SS
GC
N)
,
wh
ich
u
tili
ze
d
s
em
an
tic
in
f
o
r
m
atio
n
an
d
s
y
n
tactic
weig
h
t
m
at
r
ix
f
o
r
o
b
tain
i
n
g
tex
t
-
s
em
a
n
tic
r
ep
r
esen
tatio
n
s
.
I
n
th
is
wo
r
k
an
atten
tio
n
m
o
d
el
was
in
tr
o
d
u
ce
d
f
o
r
o
b
tain
in
g
s
p
ec
if
ic
asp
ec
t
-
h
id
d
en
co
n
tex
t
v
ec
to
r
s
.
T
h
e
SS
GC
N
ap
p
r
o
ac
h
im
p
r
o
v
e
d
tex
t
-
r
e
p
r
esen
tatio
n
s
an
d
ev
alu
ati
o
n
s
wer
e
co
n
d
u
cte
d
u
s
in
g
R
estau
r
at2
0
1
6
,
R
estau
r
an
t2
0
1
5
,
R
estau
r
an
t2
0
1
4
,
an
d
L
ap
to
p
2
0
1
4
,
wh
er
e
a
ch
iev
ed
8
9
.
7
7
%,
8
1
.
1
8
%,
8
2
.
9
6
%
,
an
d
7
5
.
8
6
%
ac
cu
r
ac
y
r
esp
ec
tiv
ely
.
J
ian
g
e
t
a
l.
[
2
4
]
,
f
o
r
s
o
lv
in
g
is
s
u
es
o
f
co
n
tex
t
-
awa
r
en
ess
,
less
ac
cu
r
ac
y
,
p
r
esen
ted
a
m
o
d
el
ca
lled
d
ee
p
c
o
n
tex
t
-
aw
ar
e
s
en
tim
en
t
an
aly
s
is
m
o
d
e
l
(
DC
ASA
M)
,
wh
ich
in
teg
r
ated
d
ee
p
Bi
L
STM
an
d
d
en
s
ely
co
n
n
e
cted
g
r
a
p
h
c
o
n
v
o
lu
ti
o
n
al
n
etwo
r
k
(
DGCN),
wh
ich
ca
p
tu
r
e
d
co
n
tex
t
u
al
v
ar
iatio
n
s
an
d
d
ep
en
d
e
n
cies.
T
h
e
d
ee
p
Bi
L
STM
ef
f
icien
tly
ca
p
tu
r
ed
s
eq
u
en
tial
d
ep
e
n
d
en
c
y
,
w
h
er
ea
s
th
e
DGCN
u
tili
ze
d
d
en
s
ely
co
n
n
ec
ted
s
tr
u
ctu
r
es
f
o
r
ca
p
tu
r
i
n
g
in
tr
icate
co
n
n
ec
tio
n
in
d
ata.
Usi
n
g
th
e
f
o
llo
win
g
,
th
e
DC
ASAM
ac
h
iev
es
h
ig
h
e
r
ac
c
u
r
ac
y
.
E
v
alu
atio
n
s
wer
e
co
n
d
u
cted
o
n
T
witter
2
0
1
4
,
L
ap
to
p
2
0
1
4
an
d
R
estau
r
an
t2
0
1
4
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
ema
n
tic
-
s
yn
ta
ctic
g
r
a
p
h
n
et
w
o
r
k
fo
r
a
s
p
ec
t
-
b
a
s
ed
s
en
timen
t a
n
a
lysi
s
(
R
ek
h
a
B
d
u
r
g
a
Ha
r
is
h
)
1817
d
atasets
,
wh
er
e
ac
h
iev
ed
7
8
.
2
7
%,
8
1
.
7
%
,
an
d
8
7
.
1
6
%
ac
cu
r
ac
y
.
B
ash
a
an
d
Sh
iv
ap
p
a
[
2
5
]
,
p
r
esen
te
d
an
ap
p
r
o
ac
h
ca
lled
m
o
d
if
ied
late
n
t
Dir
ich
let
allo
ca
tio
n
o
p
tim
ized
B
E
R
T
L
STM
f
o
r
AB
SA
(
MO
L
-
AB
SA
)
,
wh
ich
u
tili
ze
d
a
m
o
d
if
ied
-
laten
t
Dir
ich
let
-
allo
ca
tio
n
f
o
r
ex
t
r
ac
tio
n
o
f
asp
ec
ts
f
r
o
m
s
en
ten
ce
,
a
n
o
p
tim
ized
B
E
R
T
(
O‑
B
E
R
T
)
m
o
d
el
was
u
s
ed
f
o
r
s
en
tim
en
t
class
if
icatio
n
an
d
L
STM
was
u
s
ed
f
o
r
im
p
r
o
v
in
g
class
if
icatio
n
ac
cu
r
ac
y
.
E
v
alu
atio
n
s
wer
e
c
o
n
d
u
cted
o
n
Sem
E
v
al2
0
1
4
an
d
a
co
llected
h
o
s
p
ital
d
ata,
wh
er
e
f
in
d
in
g
s
s
h
o
we
d
8
5
.
2
6
%,
8
1
.
1
9
%
,
an
d
8
7
.
5
3
%
ac
cu
r
ac
y
f
o
r
R
estau
r
an
t2
0
1
4
,
L
ap
to
p
2
0
1
4
,
an
d
h
o
s
p
ital d
ata
s
et
r
esp
ec
tiv
ely
.
Hu
an
d
L
i
[
2
6
]
p
r
esen
ted
a
m
o
d
el
ca
lled
bi
-
c
h
an
n
el
g
r
a
p
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
B
I
C
-
GC
N)
f
o
r
s
o
lv
in
g
is
s
u
es
o
f
s
y
n
tactic
tr
ee
ap
p
r
o
ac
h
es
as
th
ey
in
tr
o
d
u
ce
n
o
is
e
an
d
af
f
ec
t
o
v
er
all
p
er
f
o
r
m
a
n
ce
.
T
h
e
B
I
C
-
GC
N
also
s
o
lv
ed
is
s
u
e
o
f
s
in
g
le
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
SGC
N)
ap
p
r
o
ac
h
es
as
th
ey
f
ail
to
ag
g
r
eg
ated
s
y
n
tactic
a
n
d
s
em
an
tic
s
tr
u
ctu
r
al
n
o
d
e
in
f
o
r
m
a
tio
n
,
af
f
ec
tin
g
s
en
tim
en
t
class
if
icatio
n
.
I
n
th
ei
r
wo
r
k
,
a
p
h
r
ase
-
s
tr
u
ctu
r
e
tr
ee
was
in
tr
o
d
u
ce
d
wh
ich
tr
an
s
f
o
r
m
ed
th
e
d
ata
to
s
tr
u
ctu
r
e
p
h
r
ase
m
atr
ix
.
T
h
e
ad
jace
n
t
m
atr
ix
o
f
s
tr
u
ctu
r
e
p
h
r
ase
m
atr
ix
an
d
d
ep
e
n
d
en
t
s
y
n
tactic
-
tr
ee
wer
e
m
er
g
e
d
to
f
o
r
m
in
itial
GC
N
f
o
r
en
h
an
cin
g
s
y
n
tactic
in
f
o
r
m
ati
o
n
.
T
h
e
s
em
an
tic
f
ea
tu
r
e
r
ep
r
esen
tatio
n
was
ac
h
iev
ed
u
s
in
g
MH
SA
an
d
GC
N
an
d
wer
e
th
en
f
u
s
ed
f
o
r
ac
h
iev
in
g
a
d
u
al
-
ch
an
n
el
co
m
p
lem
en
tar
y
-
lear
n
in
g
f
ea
t
u
r
e.
E
v
alu
atio
n
s
wer
e
co
n
d
u
cte
d
o
n
T
witter
2
0
1
4
,
L
ap
to
p
2
0
1
4
,
an
d
R
estau
r
an
t2
0
1
4
,
wh
e
r
e
ac
h
iev
e
d
7
8
.
2
7
%,
8
1
.
7
0
%
,
an
d
8
7
.
1
6
%
ac
cu
r
ac
y
r
esp
ec
tiv
el
y
.
Han
[
2
7
]
p
r
esen
ted
s
y
n
tacti
c
-
m
ask
ed
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
SM
-
GC
N)
ap
p
r
o
ac
h
f
o
r
u
n
d
er
s
tan
d
i
n
g
s
y
n
tactic
-
s
tr
u
ctu
r
e
in
f
o
r
m
atio
n
,
f
o
r
ef
f
ec
tiv
e
ly
p
r
e
d
ictin
g
s
en
tim
en
t
an
aly
s
is
.
T
h
e
SM
-
GC
N
ap
p
r
o
ac
h
u
tili
ze
d
s
y
n
tactic
d
ata
an
d
GC
N
f
o
r
lea
r
n
in
g
e
v
er
y
asp
ec
t
weig
h
t
i
n
s
en
ten
ce
an
d
its
r
elate
d
co
n
n
ec
tio
n
i
n
co
m
p
lete
s
en
ten
ce
.
B
y
in
co
r
p
o
r
atio
n
o
f
s
y
n
tactic
-
m
ask
m
atr
ix
,
th
e
SM
-
GC
N
p
r
o
v
id
es
k
ey
s
eg
m
en
ts
,
wh
ich
is
r
eq
u
ir
ed
f
o
r
s
en
tim
en
t
an
al
y
s
is
,
th
er
eb
y
im
p
r
o
v
in
g
s
en
tim
en
t
class
if
icatio
n
ac
cu
r
ac
y
.
Als
o
,
th
e
m
o
d
el
u
tili
ze
d
ab
s
tr
ac
t
-
g
r
ap
h
s
tr
u
ct
u
r
e
o
f
GC
N
f
o
r
in
teg
r
atin
g
n
ex
t
n
o
d
e
in
f
o
r
m
atio
n
b
y
u
tili
zin
g
tr
an
s
f
er
m
atr
ix
f
r
o
m
f
u
s
ed
s
y
n
tactic
m
ask
ed
m
atr
i
x
f
o
r
im
p
r
o
v
in
g
p
er
f
o
r
m
a
n
ce
.
E
v
alu
ati
o
n
s
wer
e
co
n
d
u
cte
d
o
n
Sem
E
v
al2
0
1
4
d
atasets
,
wh
er
e
ac
h
iev
e
d
7
7
.
2
4
%,
8
0
.
8
6
%
,
an
d
8
6
.
3
1
%
ac
c
u
r
ac
y
f
o
r
T
witter
2
0
1
4
,
L
ap
to
p
2
0
1
4
,
an
d
R
estau
r
an
t2
0
1
4
d
atasets
r
esp
ec
tiv
ely
.
T
h
e
co
m
p
lete
s
u
m
m
ar
y
o
f
liter
atu
r
e
s
u
r
v
ey
is
p
r
esen
ted
in
T
a
b
le
1
.
Fro
m
th
e
ex
ten
s
iv
e
liter
atu
r
e
s
u
r
v
ey
,
it
is
ev
id
en
t
th
at
n
u
m
e
r
o
u
s
ap
p
r
o
ac
h
es
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
f
o
r
AB
S
A
u
s
in
g
DL
an
d
g
r
ap
h
-
b
ased
m
o
d
els.
Ho
wev
er
,
a
cr
it
ical
o
b
s
er
v
atio
n
ac
r
o
s
s
m
o
s
t
o
f
th
e
wo
r
k
s
is
th
e
lim
ited
in
teg
r
atio
n
o
f
b
o
th
s
y
n
tactic
an
d
s
em
an
tic
in
f
o
r
m
atio
n
in
a
u
n
if
ie
d
m
an
n
er
.
Sev
er
al
m
o
d
els
h
a
v
e
em
p
h
asized
eith
er
s
y
n
tactic
o
r
s
em
an
tic
cu
es,
lead
in
g
to
s
u
b
o
p
tim
al
p
er
f
o
r
m
an
ce
in
co
m
p
lex
s
en
ten
ce
s
tr
u
ctu
r
es.
Fo
r
in
s
tan
ce
,
SDTG
C
N
[
2
0
]
f
o
cu
s
es
p
r
im
ar
ily
o
n
s
y
n
tactic
d
ep
en
d
en
c
y
g
r
a
p
h
s
with
o
u
t
ef
f
ec
tiv
ely
in
co
r
p
o
r
atin
g
r
ich
s
em
a
n
tic
c
o
n
tex
t.
Similar
ly
,
SAGC
N
[
2
1
]
im
p
r
o
v
es
u
p
o
n
s
y
n
tactic
f
e
atu
r
e
ex
tr
ac
tio
n
b
u
t
o
n
ly
p
ar
tially
in
te
g
r
ates
s
em
a
n
tic
k
n
o
wled
g
e
t
h
r
o
u
g
h
ex
ter
n
al
s
en
tim
en
t
r
eso
u
r
ce
s
.
W
h
ile
KHGCN
[
2
2
]
a
n
d
SS
-
GC
N
[
2
3
]
attem
p
t
to
co
m
b
in
e
m
u
ltip
le
s
o
u
r
ce
s
o
f
in
f
o
r
m
atio
n
,
th
eir
f
u
s
io
n
m
ec
h
an
is
m
s
ar
e
eith
er
s
im
p
lis
tic
o
r
lo
o
s
ely
in
teg
r
at
ed
,
r
esu
ltin
g
in
lim
ited
s
y
n
er
g
y
b
etwe
en
s
y
n
tax
an
d
s
em
a
n
tics
.
Mo
d
els
lik
e
SM
-
GC
N
[
2
7
]
a
n
d
B
I
C
-
GC
N
[
2
6
]
f
o
c
u
s
o
n
r
e
f
i
n
i
n
g
s
y
n
ta
c
t
ic
s
t
r
u
ct
u
r
e
s
b
u
t
s
t
i
ll
l
a
c
k
c
o
m
p
r
e
h
e
n
s
i
v
e
s
e
m
a
n
t
i
c
m
o
d
e
l
i
n
g
,
w
h
i
c
h
i
s
c
r
u
c
i
al
f
o
r
c
a
p
t
u
r
i
n
g
t
h
e
n
u
a
n
c
e
d
m
e
a
n
in
g
o
f
a
s
p
e
ct
t
e
r
m
s
i
n
v
a
r
y
i
n
g
c
o
n
t
e
x
t
s
.
T
o
a
d
d
r
e
s
s
t
h
i
s
g
a
p
,
t
h
is
w
o
r
k
p
r
o
p
o
s
e
d
Se
n
t
S
e
m
S
y
n
N
e
t
,
a
h
y
b
r
i
d
m
o
d
e
l
t
h
a
t
e
f
f
ec
t
i
v
e
l
y
c
o
m
b
i
n
es
d
e
ep
s
e
m
a
n
t
i
c
f
e
at
u
r
e
s
f
r
o
m
B
E
R
T
,
s
e
q
u
e
n
ti
a
l
d
e
p
e
n
d
e
n
c
i
e
s
v
i
a
B
i
L
S
T
M
,
a
n
d
s
y
n
t
ac
t
i
c
s
t
r
u
c
t
u
r
e
s
u
s
i
n
g
a
t
w
o
-
l
a
y
er
G
C
N
.
B
y
u
n
i
f
y
i
n
g
t
h
e
s
e
c
o
m
p
o
n
e
n
t
s
,
Se
n
t
S
e
m
Sy
n
N
e
t
e
n
s
u
r
e
s
r
o
b
u
s
t
as
p
e
c
t
-
s
e
n
t
i
m
e
n
t
a
l
i
g
n
m
e
n
t
a
n
d
i
m
p
r
o
v
e
s
c
l
a
s
s
i
f
ic
a
t
i
o
n
a
c
c
u
r
a
c
y
,
o
u
t
p
e
r
f
o
r
m
i
n
g
m
o
d
e
l
s
t
h
a
t
i
s
o
l
at
e
s
y
n
t
a
ct
i
c
o
r
s
e
m
a
n
ti
c
r
e
p
r
es
e
n
t
at
i
o
n
s
.
T
h
i
s
c
o
m
p
r
e
h
e
n
s
i
v
e
i
n
t
e
g
r
a
t
i
o
n
m
a
r
k
s
a
s
i
g
n
i
f
i
c
a
n
t
a
d
v
a
n
c
e
m
e
n
t
i
n
A
B
S
A
,
t
ac
k
l
in
g
l
i
m
i
t
a
t
i
o
n
s
i
d
e
n
t
i
f
i
e
d
i
n
e
x
is
t
i
n
g
w
o
r
k
s
s
u
c
h
a
s
[
2
0
]
−
[
2
2
]
,
[
2
7
]
.
T
h
e
c
o
m
p
l
e
t
e
m
e
t
h
o
d
o
f
S
e
n
t
S
e
m
S
y
n
N
e
t
i
s
d
i
s
c
u
s
s
e
d
i
n
t
h
e
n
e
x
t
s
ec
t
i
o
n
.
T
ab
le
1
.
L
iter
atu
r
e
s
u
r
v
e
y
s
u
m
m
ar
y
Ref
M
o
d
e
l
n
a
m
e
D
a
t
a
s
e
t
s
u
s
e
d
A
c
c
u
r
a
c
y
(
%)
[
1
6
]
C
P
A
-
SA
R
e
st
a
u
r
a
n
t
2
0
1
6
,
R
e
s
t
a
u
r
a
n
t
2
0
1
5
,
R
e
s
t
a
u
r
a
n
t
2
0
1
4
,
a
n
d
La
p
t
o
p
2
0
1
4
8
9
.
0
2
,
7
9
.
6
1
,
7
5
.
1
8
,
8
2
.
6
4
[
1
7
]
HGNN
R
e
st
a
u
r
a
n
t
2
0
1
6
,
R
e
s
t
a
u
r
a
n
t
2
0
1
5
,
R
e
s
t
a
u
r
a
n
t
2
0
1
4
,
La
p
t
o
p
2
0
1
4
,
a
n
d
T
w
i
t
t
e
r
2
0
1
4
8
7
.
9
2
,
8
0
.
3
7
,
7
4
.
0
8
,
8
1
.
9
1
,
8
0
.
2
9
[
1
8
]
LG
C
F
R
e
st
a
u
r
a
n
t
2
0
1
6
,
R
e
s
t
a
u
r
a
n
t
2
0
1
4
,
La
p
t
o
p
2
0
1
4
,
T
w
i
t
t
e
r
2
0
1
4
,
a
n
d
M
A
M
S
(
c
a
m
e
r
a
,
c
a
r
,
p
h
o
n
e
,
t
e
l
e
v
i
s
i
o
n
,
a
n
d
t
-
sh
i
r
t
)
9
1
.
8
7
,
8
5
.
5
2
,
8
1
.
2
9
,
7
5
.
6
9
;
M
A
M
S
:
9
7
.
2
4
,
9
8
.
2
6
,
9
7
.
5
9
,
9
1
.
6
1
,
9
3
.
8
6
[
2
0
]
S
D
TG
C
N
T
w
i
t
t
e
r
2
0
1
4
,
R
e
st
a
u
r
a
n
t
2
0
1
4
,
L
a
p
t
o
p
2
0
1
4
,
R
e
s
t
a
u
r
a
n
t
2
0
1
5
,
a
n
d
R
e
st
a
u
r
a
n
t
2
0
1
6
7
6
.
2
5
,
8
3
.
8
2
,
7
8
.
6
4
,
8
3
.
2
1
,
9
1
.
5
3
[
2
1
]
S
A
G
C
N
T
w
i
t
t
e
r
2
0
1
4
,
R
e
st
a
u
r
a
n
t
2
0
1
4
,
a
n
d
L
a
p
t
o
p
2
0
1
4
7
7
.
9
7
,
9
7
.
5
3
,
8
3
.
0
6
[
2
2
]
K
H
G
C
N
R
e
st
a
u
r
a
n
t
2
0
1
6
,
R
e
s
t
a
u
r
a
n
t
2
0
1
5
,
a
n
d
La
p
t
o
p
2
0
1
4
9
1
.
0
7
,
8
5
.
4
2
,
8
0
.
8
7
[
2
3
]
SS
-
G
C
N
R
e
st
a
u
r
a
n
t
2
0
1
6
,
R
e
s
t
a
u
r
a
n
t
2
0
1
5
,
R
e
s
t
a
u
r
a
n
t
2
0
1
4
,
a
n
d
La
p
t
o
p
2
0
1
4
8
9
.
7
7
,
8
1
.
1
8
,
8
2
.
9
6
,
7
5
.
8
6
[
2
4
]
D
C
A
S
A
M
T
w
i
t
t
e
r
2
0
1
4
,
La
p
t
o
p
2
0
1
4
,
a
n
d
R
e
s
t
a
u
r
a
n
t
2
0
1
4
7
8
.
2
7
,
8
1
.
7
0
,
8
7
.
1
6
[
2
5
]
M
O
L
-
A
B
S
A
R
e
st
a
u
r
a
n
t
2
0
1
4
,
L
a
p
t
o
p
2
0
1
4
,
a
n
d
H
o
sp
i
t
a
l
d
a
t
a
se
t
8
5
.
2
6
,
8
1
.
1
9
,
8
7
.
5
3
[
2
6
]
B
I
C
-
G
C
N
T
w
i
t
t
e
r
2
0
1
4
,
La
p
t
o
p
2
0
1
4
,
a
n
d
R
e
s
t
a
u
r
a
n
t
2
0
1
4
7
8
.
2
7
,
8
1
.
7
0
,
8
7
.
1
6
[
2
7
]
SM
-
G
C
N
T
w
i
t
t
e
r
2
0
1
4
,
La
p
t
o
p
2
0
1
4
,
a
n
d
R
e
s
t
a
u
r
a
n
t
2
0
1
4
7
7
.
2
4
,
8
0
.
8
6
,
8
6
.
3
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
8
1
4
-
1
8
2
4
1818
3.
M
E
T
H
O
D
AB
S
A
ap
p
r
o
ac
h
es
s
ee
k
t
o
d
eter
m
in
e
s
en
tim
en
t
ex
p
r
ess
ed
t
o
war
d
a
s
p
ec
if
ic
asp
ec
t
o
f
a
n
en
tity
in
a
s
en
ten
ce
.
Fo
r
e
x
am
p
le,
in
r
e
v
iew
“T
h
e
b
atter
y
life
o
f
th
is
p
h
o
n
e
is
am
az
in
g
,
b
u
t
th
e
c
am
er
a
is
p
o
o
r
,
”
th
e
s
en
tim
en
ts
to
war
d
“
b
atter
y
lif
e”
an
d
“c
am
er
a”
ar
e
d
if
f
e
r
en
t.
Hen
ce
,
th
er
e
is
a
r
eq
u
i
r
em
en
t
f
o
r
a
g
o
o
d
AB
SA
m
o
d
el
wh
ich
ca
p
tu
r
es
b
o
th
s
em
an
tic
co
n
tex
t
an
d
s
y
n
tactic
r
elatio
n
s
h
ip
s
f
r
o
m
th
e
s
en
ten
ce
p
r
o
v
i
d
in
g
b
etter
AB
S
A
p
r
ed
ictio
n
.
Hen
ce
,
th
is
wo
r
k
p
r
esen
ts
a
h
y
b
r
id
DL
a
r
ch
itectu
r
e
f
o
r
AB
SA,
ca
lled
as
Sen
tSem
Sy
n
Net
in
s
p
ir
ed
b
y
w
o
r
k
s
p
r
esen
ted
b
y
Dev
lin
et
a
l.
[
2
8
]
.
3
.
1
.
Arc
hite
ct
ure
F
i
g
u
r
e
1
p
r
e
s
e
n
t
s
S
e
n
t
S
em
S
y
n
N
e
t
a
r
ch
i
t
e
c
tu
r
e,
w
h
er
e
f
ir
s
t
i
t
c
o
n
s
i
d
er
s
i
n
p
u
t
d
a
t
a
,
wh
i
ch
c
o
n
s
i
s
t
s
o
f
r
ev
i
e
w
s
e
n
t
en
c
e
,
s
p
e
c
i
f
i
c
a
s
p
e
c
t
t
e
r
m
an
d
s
en
t
i
m
en
t
lab
e
l
.
T
h
e
f
ir
s
t
s
t
ag
e
o
f
S
en
t
Se
m
S
y
n
N
e
t
in
c
l
u
d
e
s
i
n
p
u
t
r
e
p
r
e
s
e
n
ta
t
i
o
n
,
w
h
er
e
B
E
R
T
i
s
u
s
e
d
f
o
r
c
o
n
v
er
t
i
n
g
t
e
x
t
i
n
p
u
t
in
t
o
r
ic
h
co
n
t
ex
t
u
a
l
em
b
ed
d
in
g
,
e
f
f
e
c
t
i
v
e
l
y
ca
p
tu
r
i
n
g
s
e
m
an
t
ic
i
n
f
o
r
m
a
t
i
o
n
[
2
9
]
.
T
h
e
s
e
e
m
b
e
d
d
i
n
g
s
a
r
e
th
en
p
a
s
s
e
d
t
o
B
i
L
S
T
M
,
a
l
l
o
w
in
g
S
e
n
t
S
e
m
S
y
n
N
e
t
m
o
d
e
l
to
u
n
d
e
r
s
t
an
d
s
e
q
u
e
n
t
ia
l
w
o
r
d
r
e
l
at
i
o
n
s
h
ip
s
a
n
d
p
o
s
i
t
io
n
a
l
in
f
o
r
m
a
t
i
o
n
e
f
f
e
c
t
iv
e
l
y
.
F
u
r
t
h
e
r
,
g
r
am
m
a
t
i
ca
l
c
o
n
t
e
x
t
u
n
d
e
r
s
t
a
n
d
i
n
g
i
s
h
a
n
d
l
ed
t
h
r
o
u
g
h
g
r
a
p
h
c
o
n
s
t
r
u
c
t
i
o
n
,
w
h
er
e
s
y
n
ta
c
t
i
c
d
e
p
en
d
e
n
cy
g
r
a
p
h
o
f
s
en
te
n
c
e
i
s
c
r
ea
t
e
d
[
3
0
]
.
T
h
i
s
g
r
ap
h
i
s
t
h
en
f
ed
t
o
t
wo
-
l
ay
e
r
G
C
N
f
o
r
s
y
n
t
a
c
t
i
c
-
i
n
f
o
r
m
a
t
io
n
ex
t
r
ac
ti
o
n
,
a
l
l
o
w
in
g
S
en
t
S
e
m
S
y
n
N
e
t
t
o
l
e
ar
n
f
r
o
m
g
r
a
m
m
at
i
c
a
l
s
t
r
u
c
tu
r
e
a
n
d
l
o
n
g
-
r
a
n
g
e
d
e
p
en
d
en
c
i
e
s
in
s
e
n
t
e
n
c
e.
T
h
e
o
u
t
p
u
t
i
s
th
en
p
a
s
s
ed
t
o
a
n
a
s
p
e
c
t
-
a
w
ar
e
a
g
g
r
eg
a
t
i
o
n
an
d
p
o
o
li
n
g
s
t
a
g
e,
w
h
er
e
m
e
a
n
-
p
o
o
l
in
g
i
s
u
s
ed
to
g
en
e
r
a
t
e
f
ix
e
d
-
l
en
g
th
v
e
c
t
o
r
r
ep
r
e
s
en
t
i
n
g
s
e
n
t
i
m
en
t
t
o
w
ar
d
a
s
p
e
c
t
[
3
1
]
.
T
h
i
s
v
e
c
to
r
i
s
t
h
e
n
u
s
ed
i
n
s
en
t
i
m
en
t
c
l
a
s
s
i
f
ic
a
t
i
o
n
,
w
h
er
e
S
o
f
t
M
ax
l
a
y
e
r
p
r
ed
i
c
t
s
s
e
n
t
i
m
e
n
t
.
F
i
n
a
l
l
y
,
S
en
t
S
e
m
S
y
n
N
e
t
p
e
r
f
o
r
m
an
c
e
i
s
e
v
a
l
u
a
t
ed
u
s
i
n
g
s
ta
n
d
ar
d
p
er
f
o
r
m
a
n
c
e
m
e
tr
i
c
s
u
s
e
d
f
o
r
A
B
S
A
.
Fig
u
r
e
1
.
Sen
tSem
Sy
n
Net
ar
c
h
itectu
r
e
3
.
2
.
I
np
ut
re
presenta
t
io
n
C
o
n
s
id
er
an
in
p
u
t
f
r
o
m
d
ata
s
et
wh
ich
co
n
s
is
ts
o
f
s
en
ten
ce
wh
ich
h
as
an
asp
ec
t
ter
m
at
a
p
o
s
itio
n
,
s
u
ch
th
at
,
∈
.
Als
o
co
n
s
id
er
a
lab
el
f
o
r
e
v
er
y
s
en
ten
ce
d
en
o
ted
as
∈
{
0
,
1
,
2
}
,
wh
ich
r
ep
r
esen
ts
s
en
tim
en
ts
f
o
r
s
en
ten
ce
as
n
eg
ativ
e,
n
eu
tr
al
an
d
p
o
s
itiv
e
r
esp
ec
tiv
ely
.
As
s
en
t
en
ce
co
n
s
is
ts
o
f
m
u
ltip
le
wo
r
d
s
,
let
th
e
wo
r
d
s
b
e
r
ep
r
esen
ted
as
1
,
2
,
3
,
…
,
.
I
n
th
is
wo
r
k
,
th
e
B
E
R
T
m
o
d
el
h
as
b
ee
n
u
s
ed
to
em
b
ed
in
p
u
t
s
en
ten
ce
an
d
asp
ec
t
in
a
way
s
u
ch
t
h
at
it
ca
p
t
u
r
es
r
ich
co
n
tex
tu
a
l
r
elev
an
ce
an
d
s
em
an
tic
in
f
o
r
m
atio
n
o
f
ea
ch
wo
r
d
;
h
e
n
ce
,
th
e
f
ir
s
t
s
tep
o
f
th
e
B
E
R
T
ap
p
r
o
ac
h
is
to
en
c
o
d
e
th
e
s
en
ten
ce
.
C
o
n
s
id
er
s
en
ten
ce
as
=
{
1
,
2
,
3
,
…
,
}
,
wh
ich
m
ay
h
av
e
o
n
e
o
r
m
o
r
e
asp
e
cts
b
ein
g
r
ep
r
esen
ted
as
=
[
1
,
2
,
…
,
]
.
T
h
e
in
p
u
t seq
u
en
ce
f
e
d
in
to
t
h
e
B
E
R
T
m
o
d
el
is
r
ep
r
esen
ted
b
y
(
1
)
.
=
[
[
]
,
1
,
2
,
3
,
…
,
,
[
]
,
1
,
2
,
…
,
,
[
]
]
(1
)
I
n
(
1
)
,
[
]
an
d
[
]
d
en
o
tes
B
E
R
T
to
k
en
s
,
wh
ich
r
ep
r
esen
t
class
if
i
er
to
k
en
a
n
d
s
ep
ar
ato
r
t
o
k
en
.
T
h
e
(
1
)
d
ef
in
es
h
o
w
s
en
ten
ce
an
d
asp
ec
t
ar
e
co
m
b
in
e
as
in
p
u
t
u
s
in
g
B
E
R
T
[
]
an
d
[
]
to
k
en
s
,
h
elp
in
g
m
o
d
el
co
n
d
itio
n
m
e
an
in
g
o
f
s
en
ten
ce
o
n
asp
ec
t.
T
h
e
in
p
u
t
s
en
ten
ce
r
e
p
r
ese
n
ted
by
(
1
)
is
th
en
en
co
d
ed
u
s
in
g
B
E
R
T
m
o
d
el
Dev
lin
et
a
l.
[
2
8
]
as p
r
esen
ted
i
n
(
2
)
.
0
=
(
)
∈
ℝ
×
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
ema
n
tic
-
s
yn
ta
ctic
g
r
a
p
h
n
et
w
o
r
k
fo
r
a
s
p
ec
t
-
b
a
s
ed
s
en
timen
t a
n
a
lysi
s
(
R
ek
h
a
B
d
u
r
g
a
Ha
r
is
h
)
1819
I
n
(
2
)
,
d
en
o
tes
m
ax
im
u
m
s
e
q
u
en
ce
len
g
t
h
,
d
en
o
tes
h
id
d
en
s
ize
f
o
r
B
E
R
T
m
o
d
el,
wh
er
e
th
is
wo
r
k
h
as
c
o
n
s
id
er
ed
=
768
,
0
=
ℎ
1
0
,
ℎ
2
0
,
…
ℎ
0
d
en
o
tes
co
n
tex
tu
alize
d
em
b
ed
d
in
g
o
f
to
k
en
.
I
n
th
is
wo
r
k
,
th
e
B
E
R
T
lev
er
ag
es
m
u
lti
-
h
ea
d
atten
tio
n
an
d
p
o
s
itio
n
al
-
en
co
d
in
g
f
o
r
e
n
co
d
in
g
d
ee
p
-
s
em
an
tic
f
ea
tu
r
es.
T
h
is
lay
s
s
em
an
tic
f
o
u
n
d
atio
n
f
o
r
Sen
tSem
Sy
n
Net
ap
p
r
o
a
ch
,
en
s
u
r
in
g
asp
ec
t
-
awa
r
e
u
n
d
er
s
tan
d
in
g
.
I
n
th
e
B
E
R
T
m
o
d
el,
it
is
s
ee
n
th
at
t
h
e
m
o
d
el
f
ails
to
ca
p
tu
r
e
tem
p
o
r
al
d
ep
e
n
d
en
cies
an
d
wo
r
d
-
o
r
d
er
,
h
en
ce
,
in
th
is
wo
r
k
B
iLST
M
ap
p
r
o
ac
h
h
as b
ee
n
u
tili
ze
d
to
s
o
lv
e
t
h
is
p
r
o
b
l
em
.
3
.
2
.
1
.
Sequ
ent
ia
l
m
o
dellin
g
us
ing
B
i
L
S
T
M
I
n
NL
P
,
s
eq
u
en
tial
u
n
d
e
r
s
tan
d
in
g
o
f
s
en
ten
ce
s
is
o
f
ten
a
ch
iev
ed
u
s
in
g
L
STM
n
etwo
r
k
s
,
wh
ich
ef
f
icien
tly
ca
p
tu
r
es
tem
p
o
r
al
d
ep
en
d
e
n
cies.
Ho
wev
er
,
s
tan
d
ar
d
L
STM
m
o
d
els
ar
e
lim
ited
in
ca
p
tu
r
in
g
wo
r
d
o
r
d
er
f
r
o
m
b
o
th
p
ast
an
d
f
u
tu
r
e
co
n
te
x
ts
.
T
h
er
ef
o
r
e,
t
h
is
wo
r
k
e
m
p
lo
y
s
a
B
iLST
M
m
o
d
el
to
b
etter
ca
p
tu
r
e
b
o
th
tem
p
o
r
al
d
ep
en
d
en
cies
an
d
co
n
tex
tu
al
wo
r
d
o
r
d
er
.
I
n
B
iLST
M,
t
h
e
o
u
tp
u
t
f
r
o
m
B
E
R
T
m
o
d
el,
i.e
.
,
0
i
s
co
n
s
id
er
ed
as
in
p
u
t
f
o
r
r
ein
f
o
r
cin
g
wo
r
d
-
o
r
d
er
i
n
f
o
r
m
atio
n
an
d
ca
p
t
u
r
in
g
tem
p
o
r
al
d
e
p
en
d
en
cies
in
b
o
th
d
ir
ec
tio
n
s
;
h
en
ce
,
th
is
wo
r
k
co
m
p
u
tes
f
o
r
war
d
an
d
b
ac
k
war
d
L
STM
h
id
d
en
s
tates,
m
o
d
ellin
g
an
d
u
n
d
er
s
tan
d
i
n
g
s
en
ten
ce
f
lo
w
i
n
b
o
th
d
ir
ec
tio
n
s
.
T
h
e
f
o
r
war
d
an
d
b
ac
k
war
d
L
STM
h
id
d
en
s
tates a
r
e
ev
alu
ated
u
s
in
g
(
3
)
an
d
(
4
)
r
esp
ec
tiv
ely
.
ℎ
⃗
⃗
⃗
,
⃗
⃗
⃗
=
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
(
ℎ
0
,
−
1
⃗
⃗
⃗
⃗
⃗
⃗
⃗
)
(
3
)
ℎ
⃖
⃗
⃗
⃗
,
⃖
⃗
⃗
⃗
=
⃖
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
(
ℎ
0
,
+
1
⃖
⃗
⃗
⃗
⃗
⃗
⃗
⃗
)
(
4
)
I
n
(
3
)
,
ℎ
⃗
⃗
⃗
d
en
o
tes
h
id
d
e
n
-
s
tate
at
tim
e
-
s
tep
an
d
⃗
⃗
⃗
d
en
o
tes
ce
ll
-
s
tate
at
in
f
o
r
war
d
d
ir
ec
tio
n
.
T
h
e
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
d
en
o
tes
u
n
it
p
r
o
ce
s
s
in
g
s
eq
u
en
ce
f
r
o
m
lef
t
to
r
i
g
h
t,
i.e
.
,
f
r
o
m
p
ast
to
f
u
t
u
r
e,
ℎ
0
d
e
n
o
tes
i
n
p
u
t
wo
r
d
em
b
ed
d
in
g
at
,
wh
er
e
ℎ
0
∈
0
an
d
−
1
⃗
⃗
⃗
⃗
⃗
⃗
⃗
d
en
o
tes
ce
ll
-
s
tate
f
r
o
m
p
r
ev
io
u
s
tim
e
-
s
tep
in
f
o
r
war
d
d
ir
ec
t
io
n
.
I
n
(
4
)
,
ℎ
⃖
⃗
⃗
⃗
d
en
o
tes
h
id
d
en
-
s
tate
at
an
d
⃖
⃗
⃗
⃗
d
en
o
tes
ce
ll
-
s
tate
at
in
b
ac
k
war
d
d
ir
ec
tio
n
.
T
h
e
⃖
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
d
en
o
tes
u
n
it
p
r
o
ce
s
s
in
g
s
eq
u
en
ce
f
r
o
m
r
ig
h
t
to
lef
t,
i.e
.
,
f
r
o
m
f
u
t
u
r
e
to
p
ast
an
d
+
1
⃖
⃗
⃗
⃗
⃗
⃗
⃗
⃗
d
en
o
tes
ce
ll
-
s
tate
f
r
o
m
n
ex
t
tim
e
-
s
tep
in
b
ac
k
war
d
d
ir
ec
ti
o
n
.
Usi
n
g
th
e
f
o
r
war
d
an
d
b
a
ck
war
d
L
STM
h
id
d
en
s
tates,
th
e
f
in
al
B
iLST
M
o
u
tp
u
t is ac
h
iev
e
d
b
y
co
m
b
i
n
in
g
b
o
t
h
th
ese
s
tates,
wh
ich
is
r
ep
r
esen
ted
u
s
in
g
(
5
)
.
ℎ
1
=
[
ℎ
⃗
⃗
⃗
,
ℎ
⃖
⃗
⃗
⃗
]
∈
ℝ
2ℎ
(
5
)
I
n
(
5
)
,
ℎ
1
∈
1
∈
ℝ
×
2ℎ
,
wh
er
e
ℎ
d
en
o
tes
h
id
d
e
n
s
tate
o
f
ea
ch
L
STM
d
ir
ec
tio
n
.
T
h
e
f
in
al
o
u
tp
u
t
o
f
th
e
B
iLST
M
m
o
d
el
r
etain
s
s
eq
u
en
tial
co
n
tex
t
an
d
r
ic
h
lo
ca
l
co
n
tex
ts
,
wh
ich
en
h
a
n
ce
s
th
e
s
eq
u
en
tial
co
h
er
en
ce
,
im
p
o
r
ta
n
t
f
o
r
p
h
r
ases
lik
e
“o
n
ly
s
lig
h
tly
b
etter
”
o
r
“n
o
t
v
er
y
g
o
o
d
”.
Af
ter
ex
tr
ac
tio
n
o
f
th
e
s
eq
u
en
tial
co
n
tex
t,
it
is
im
p
o
r
tan
t
to
m
o
d
el
g
r
a
m
m
atic
al
s
tr
u
ctu
r
e,
h
en
ce
,
in
th
is
wo
r
k
,
a
s
y
n
tactic
d
ep
en
d
e
n
cy
g
r
ap
h
is
b
u
ilt,
wh
ich
is
d
is
cu
s
s
ed
in
th
e
n
ex
t sec
tio
n
.
3
.
2
.
2
.
Dependency
g
ra
ph
co
ns
t
ruct
io
n
W
h
ile
B
iLST
M
ef
f
ec
tiv
ely
c
ap
tu
r
es
s
eq
u
en
tial
co
n
tex
t
o
f
s
en
ten
ce
,
u
n
d
er
s
tan
d
in
g
g
r
a
m
m
atica
l
s
tr
u
ctu
r
e
r
eq
u
ir
es
d
ee
p
e
r
s
y
n
tactic
in
f
o
r
m
atio
n
.
T
h
er
ef
o
r
e,
th
is
wo
r
k
in
co
r
p
o
r
ates
a
s
y
n
tactic
d
ep
en
d
en
c
y
g
r
ap
h
to
m
o
d
el
th
e
g
r
am
m
at
ical
r
elatio
n
s
h
ip
s
b
etwe
en
w
o
r
d
s
.
Fo
r
co
n
s
tr
u
ctio
n
o
f
s
y
n
tactic
d
ep
en
d
en
cy
g
r
ap
h
,
ev
er
y
to
k
en
f
r
o
m
s
en
ten
ce
is
co
n
s
id
er
ed
as
n
o
d
e,
an
d
ed
g
es
r
ef
lect
d
e
p
en
d
e
n
cy
r
e
latio
n
s
h
ip
wh
ic
h
wer
e
p
ar
s
ed
u
s
in
g
b
u
ilt
-
in
lib
r
ar
y
,
Sp
aCy
.
C
o
n
s
id
er
g
r
ap
h
=
{
,
}
,
wh
er
e
=
{
1
,
2
,
…
,
}
r
ep
r
esen
t
to
k
en
s
,
d
en
o
tes
ed
g
es
an
d
(
,
)
∈
o
n
ly
if
a
d
ep
en
d
en
cy
r
elatio
n
b
etwe
en
an
d
ex
is
ts
.
Hen
ce
,
f
r
o
m
th
is
th
e
a
d
jace
n
cy
m
atr
ix
is
co
n
s
tr
u
cted
as
p
r
esen
ted
i
n
(
6
)
,
w
h
ich
ca
p
tu
r
es
r
elatio
n
s
h
ip
s
b
etwe
en
wo
r
d
s
in
d
ep
en
d
en
cy
p
ar
s
e.
I
n
th
is
wo
r
k
,
n
o
r
m
aliza
tio
n
is
ap
p
l
ied
to
en
s
u
r
e
n
u
m
er
ical
s
tab
ilit
y
,
f
o
llo
win
g
th
e
ap
p
r
o
ac
h
p
r
o
p
o
s
ed
in
t
h
e
G
C
N
b
y
Kip
f
a
n
d
W
ellin
g
[
2
9
]
.
Fu
r
th
er
,
in
th
e
d
ep
en
d
en
cy
g
r
ap
h
,
th
e
d
eg
r
ee
m
atr
ix
is
ev
alu
ated
u
s
in
g
(
7
)
.
I
n
(
7
)
,
d
en
o
tes
id
en
tity
m
atr
ix
.
Fu
r
th
er
,
f
o
r
n
o
r
m
alizin
g
ad
jace
n
c
y
m
atr
ix
f
o
r
s
tab
le
g
r
ap
h
co
m
p
u
tatio
n
s
,
a
s
y
m
m
etr
ic
n
o
r
m
aliza
tio
n
h
as
b
ee
n
ap
p
lied
wh
ich
is
r
ep
r
esen
ted
u
s
in
g
(
8
)
.
Usi
n
g
th
e
(
8
)
,
th
e
f
in
al
s
y
n
tactic
g
r
ap
h
h
a
v
in
g
wo
r
d
-
in
f
o
r
m
atio
n
,
tem
p
o
r
al
d
ep
en
d
en
c
y
a
n
d
g
r
am
m
atica
l
c
o
n
tex
t
is
ac
h
iev
ed
.
T
h
e
s
y
n
tactic
g
r
ap
h
ac
h
iev
e
d
u
s
in
g
(
8
)
is
cr
itical
in
r
ec
o
g
n
izin
g
lo
n
g
-
d
is
tan
ce
g
r
am
m
at
ical
r
elatio
n
s
,
i.e
.
,
co
n
n
ec
tin
g
an
asp
ec
t
with
it
s
r
esp
ec
tiv
e
m
o
d
if
ie
r
wo
r
d
.
Mo
r
eo
v
er
,
th
e
(
8
)
en
s
u
r
es
th
at
in
f
o
r
m
atio
n
is
ag
g
r
eg
ated
an
d
s
ca
led
f
air
ly
ac
r
o
s
s
v
ar
y
in
g
n
o
d
e
d
eg
r
ee
s
.
Fu
r
th
er
,
th
e
co
n
s
tr
u
cted
g
r
a
p
h
is
p
ass
ed
o
n
to
th
e
p
r
o
p
o
s
ed
GC
N
ap
p
r
o
ac
h
.
=
{
1
=
(
,
)
∈
0
ℎ
(
6
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
8
1
4
-
1
8
2
4
1820
̃
=
∑
(
+
)
(
7
)
̂
=
̃
−
1
2
(
+
)
̃
−
1
2
(
8
)
3
.
2
.
3
.
G
ra
ph
co
nv
o
lutio
na
l n
et
wo
rk
I
n
th
is
wo
r
k
,
in
s
tead
o
f
u
s
i
n
g
o
n
e
la
y
er
o
f
GC
N,
th
is
wo
r
k
h
as
ap
p
lied
two
GC
N
lay
er
s
f
o
r
p
r
o
p
a
g
atin
g
s
y
n
tactic
i
n
f
o
r
m
atio
n
ac
r
o
s
s
g
r
a
p
h
s
u
s
in
g
o
u
tp
u
ts
f
r
o
m
B
iLST
M
an
d
d
ep
en
d
e
n
cy
g
r
ap
h
co
n
s
tr
u
ctio
n
as
in
itial
n
o
d
e
f
ea
tu
r
es,
wh
ich
th
er
eb
y
allo
ws
ev
er
y
wo
r
d
to
lear
n
f
r
o
m
its
s
y
n
tactic
n
eig
h
b
o
r
s
.
T
h
e
f
ir
s
t
lay
er
o
f
GC
N
is
ev
alu
ated
u
s
in
g
(
9
)
.
I
n
(
9
)
,
d
en
o
tes
th
e
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
a
n
d
(
1
)
∈
ℝ
2ℎ
×
ℎ
d
en
o
tes
tr
ain
ab
le
weig
h
ts
.
Fu
r
th
er
,
t
h
e
s
ec
o
n
d
lay
er
o
f
GC
N
co
n
s
id
er
s
in
p
u
t
o
f
f
ir
s
t
lay
er
,
wh
ich
d
e
ep
en
s
g
r
a
p
h
f
ea
tu
r
e
lear
n
in
g
,
p
r
o
d
u
cin
g
f
in
al
n
o
d
e
r
ep
r
esen
tatio
n
s
as
3
.
T
h
e
s
ec
o
n
d
lay
er
o
f
GC
N
is
ev
alu
ated
u
s
in
g
(
1
0
)
.
2
=
(
̂
∙
1
∙
1
)
(
9
)
3
=
(
̂
∙
2
∙
2
)
(
1
0
)
I
n
(
1
0
)
,
2
d
en
o
tes
tr
ain
ab
le
weig
h
t
m
atr
ix
f
o
r
s
ec
o
n
d
lay
er
o
f
GC
N
an
d
2
∈
ℝ
ℎ
×
ℎ
.
Usi
n
g
th
e
f
o
llo
win
g
GC
N
ap
p
r
o
ac
h
,
allo
wed
ea
ch
wo
r
d
to
in
c
o
r
p
o
r
ate
in
f
o
r
m
atio
n
f
r
o
m
s
y
n
tactica
lly
co
n
n
ec
ted
wo
r
d
s
,
m
o
d
elin
g
l
o
n
g
-
r
an
g
e
d
ep
e
n
d
e
n
cy
im
p
o
r
tan
t
f
o
r
AB
SA.
Al
s
o
,
th
e
two
lay
e
r
GC
N
ap
p
r
o
a
ch
m
ad
e
th
e
m
o
d
e
l
s
tr
u
ctu
r
e
-
awa
r
e,
i
m
p
r
o
v
in
g
u
n
d
er
s
tan
d
in
g
o
f
h
o
w
s
en
tim
en
t
ter
m
s
ar
e
r
elate
d
to
asp
ec
ts
s
y
n
tactica
lly
.
Fu
r
th
er
,
f
o
r
p
r
e
v
en
tin
g
o
v
er
f
itti
n
g
,
a
d
r
o
p
o
u
t
lay
er
was
ap
p
lied
af
ter
B
iLST
M
an
d
GC
N
lay
er
s
,
wh
ich
r
an
d
o
m
l
y
ze
r
o
ed
-
o
u
t
f
ea
t
u
r
es
d
u
r
in
g
tr
ai
n
in
g
.
Fu
r
t
h
er
,
th
e
m
atr
ix
ac
h
iev
ed
u
s
in
g
s
ec
o
n
d
lay
er
o
f
GC
N
was
f
u
r
th
er
p
ass
ed
o
n
to
ag
g
r
e
g
ati
o
n
an
d
p
o
o
lin
g
lay
e
r
wh
ich
f
u
r
th
er
class
if
ied
th
e
s
en
tim
en
ts
.
3
.
2
.
4
.
Asp
ec
t
-
a
wa
re
a
g
g
re
g
a
t
io
n,
po
o
lin
g
a
nd
s
ent
im
ent
cla
s
s
if
ica
t
io
n
Fo
r
d
er
iv
i
n
g
f
ix
ed
-
le
n
g
th
s
en
ten
ce
r
e
p
r
esen
tatio
n
f
r
o
m
3
,
th
is
wo
r
k
p
er
f
o
r
m
e
d
asp
ec
t
-
awa
r
e
ag
g
r
eg
atio
n
an
d
p
o
o
lin
g
o
v
er
all
to
k
en
em
b
ed
d
in
g
s
f
r
o
m
th
e
two
-
lay
er
GC
N
ap
p
r
o
ac
h
.
A
f
ter
th
e
ag
g
r
e
g
atio
n
o
f
all
th
e
o
u
t
p
u
ts
f
r
o
m
t
h
e
3
,
a
m
ea
n
p
o
o
lin
g
was
ap
p
lied
u
s
in
g
(
1
1
)
.
I
n
(
1
1
)
,
d
en
o
tes
lo
s
s
f
u
n
ctio
n
an
d
ℎ
3
∈
3
.
B
y
u
s
in
g
th
e
m
ea
n
-
p
o
o
lin
g
o
v
er
all
n
o
d
e
em
b
ed
d
in
g
s
f
r
o
m
GC
N
o
u
tp
u
t,
t
h
is
s
tep
co
m
p
r
ess
es
all
r
elev
an
t
f
ea
tu
r
es
in
to
o
n
e
-
v
e
cto
r
,
en
ab
lin
g
class
if
icatio
n
.
T
h
e
m
ain
aim
o
f
class
if
icati
o
n
was
to
p
r
ed
ict
s
en
tim
en
t
class
,
i.e
.
,
n
eu
tr
al
o
r
p
o
s
itiv
e
o
r
n
eg
ativ
e
co
n
s
id
er
in
g
asp
ec
t.
Fo
r
class
if
icatio
n
,
th
e
p
o
o
led
-
v
ec
to
r
was
p
ass
ed
th
r
o
u
g
h
f
u
lly
-
c
o
n
n
ec
ted
lay
er
(
FC
L
)
wh
ich
f
o
llo
wed
So
f
tMa
x
f
o
r
s
en
tim
en
t
class
if
ica
tio
n
.
I
n
FC
L
,
th
e
p
o
o
led
-
v
ec
to
r
was c
o
n
v
er
ted
to
class
lo
g
its
f
o
r
cla
s
s
if
icatio
n
u
s
in
g
(
1
2
)
.
=
1
ℒ
∑
ℎ
3
∈
ℎ
ℒ
=
1
(
1
1
)
=
∙
+
∈
ℝ
(
1
2
)
I
n
(
1
2
)
,
d
en
o
tes
weig
h
t
o
f
FC
L
,
d
en
o
tes
b
ias
-
ter
m
f
o
r
FC
L
an
d
d
en
o
tes
class
if
icatio
n
class
,
i.e
.
,
n
eu
tr
al
o
r
p
o
s
itiv
e
o
r
n
eg
ativ
e
.
Fu
r
th
er
,
th
e
lo
g
its
was
p
ass
ed
o
n
to
So
f
tMa
x
,
wh
er
e
s
o
f
t
m
a
x
co
n
v
er
s
lo
g
its
in
to
class
if
icatio
n
p
r
o
b
ab
ilit
ies.
T
h
e
So
f
tMa
x
f
u
n
ctio
n
is
p
e
r
f
o
r
m
ed
u
s
in
g
(
1
3
)
.
I
n
(
1
3
)
,
̂
d
en
o
tes
p
r
ed
icte
d
p
r
o
b
ab
ilit
y
o
f
s
en
tim
en
t
class
if
icatio
n
.
Fin
ally
u
s
in
g
th
e
So
f
tMa
x
f
u
n
ctio
n
,
th
e
f
in
a
l
s
en
tim
en
t
p
r
ed
ictio
n
is
p
er
f
o
r
m
ed
u
s
in
g
(
1
4
)
.
I
n
(
1
4
)
,
̂
d
en
o
tes
p
r
ed
icted
s
en
tim
en
t.
I
n
(
1
4
)
,
th
e
f
in
al
p
r
ed
ictio
n
is
m
ad
e
u
s
in
g
o
v
er
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
.
Fu
r
t
h
er
,
to
h
an
d
le
class
im
b
alan
ce
is
s
u
es,
a
weig
h
ted
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
w
as u
s
ed
in
(
1
1
)
an
d
d
u
r
in
g
tr
ai
n
in
g
.
T
h
e
lo
s
s
f
u
n
ctio
n
was e
v
alu
ated
u
s
in
g
(
1
5
)
.
̂
=
(
)
=
ex
p
(
)
∑
ex
p
(
)
=
1
(
1
3
)
̂
=
(
̂
)
(
1
4
)
ℒ
=
−
∑
[
=
1
]
∙
l
og
(
=
1
̂
)
(
1
5
)
←
−
∙
(
1
6
)
I
n
(
1
5
)
,
d
en
o
tes
in
v
er
s
e
-
f
r
eq
u
en
cy
class
weig
h
ts
an
d
d
en
o
tes
ac
tu
al
s
en
tim
en
t
class
.
T
h
e
lo
s
s
f
u
n
ctio
n
ℒ
ass
ig
n
s
h
ig
h
er
weig
h
ts
to
u
n
d
er
r
e
p
r
esen
ted
class
es.
Fo
r
h
an
d
lin
g
th
e
weig
h
t
o
p
tim
izatio
n
s
in
th
e
m
o
d
el,
th
is
wo
r
k
u
s
es
Ad
am
o
p
tim
izer
with
weig
h
t d
ec
a
y
d
u
r
in
g
tr
ain
in
g
wh
ich
is
co
m
p
u
te
d
u
s
in
g
(
1
6
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
ema
n
tic
-
s
yn
ta
ctic
g
r
a
p
h
n
et
w
o
r
k
fo
r
a
s
p
ec
t
-
b
a
s
ed
s
en
timen
t a
n
a
lysi
s
(
R
ek
h
a
B
d
u
r
g
a
Ha
r
is
h
)
1821
I
n
(
1
6
)
,
in
clu
d
es
all
m
o
d
el
p
ar
am
eter
s
.
Fo
r
m
th
e
ab
o
v
e
ap
p
r
o
ac
h
,
it
is
s
ee
n
th
at
th
e
Sen
tSem
Sy
n
Net
ap
p
r
o
ac
h
le
v
er
ag
es
s
em
an
tic
d
ep
th
o
f
B
E
R
T
f
o
r
r
ep
r
esen
tin
g
in
p
u
t
d
ata,
s
eq
u
en
tial
m
o
d
elin
g
o
f
B
iLST
M
f
o
r
u
n
d
er
s
tan
d
in
g
wo
r
d
-
o
r
d
er
a
n
d
te
m
p
o
r
al
d
e
p
en
d
e
n
cy
e
x
tr
ac
tio
n
,
an
d
th
e
s
y
n
tactic
p
r
ec
is
io
n
o
f
GC
Ns
to
p
r
ed
ict
asp
ec
t
-
lev
el
s
en
tim
en
t
ef
f
ec
tiv
ely
.
B
y
in
teg
r
atin
g
s
y
n
tax
th
r
o
u
g
h
d
ep
e
n
d
en
c
y
g
r
ap
h
s
,
it
ca
p
tu
r
es
r
elatio
n
s
h
ip
s
th
at
ar
e
o
f
ten
m
is
s
ed
b
y
att
en
tio
n
alo
n
e.
I
n
th
e
n
e
x
t
s
ec
tio
n
,
th
e
r
esu
lts
o
f
th
e
Sen
tSem
Sy
n
Net
m
o
d
el
ar
e
d
is
cu
s
s
ed
in
d
etail.
4.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
Fo
r
th
is
s
tu
d
y
,
th
e
Sem
E
v
al2
0
1
4
T
ask
4
d
ataset
was
c
o
n
s
id
e
r
ed
.
T
h
e
Sem
E
v
al
2
0
1
4
T
ask
4
d
ataset
is
a
wid
ely
u
s
ed
b
en
c
h
m
ar
k
f
o
r
AB
SA.
I
t
co
n
s
is
ts
o
f
cu
s
to
m
er
r
ev
iews
f
r
o
m
two
d
o
m
a
in
s
:
r
estau
r
an
t
an
d
lap
to
p
.
E
ac
h
r
ev
iew
is
an
n
o
tat
ed
with
o
n
e
o
r
m
o
r
e
asp
ec
t
ter
m
s
an
d
th
eir
c
o
r
r
esp
o
n
d
in
g
s
en
tim
en
t
p
o
lar
ities
(
p
o
s
itiv
e,
n
eg
ativ
e,
o
r
n
e
u
tr
al
)
.
T
h
e
d
ataset
in
clu
d
es
lab
ele
d
asp
ec
t
ca
teg
o
r
ies
(
e.
g
.
,
f
o
o
d
,
s
er
v
ice,
p
r
ice
f
o
r
r
estau
r
an
ts
;
b
atter
y
,
d
is
p
lay
,
an
d
p
e
r
f
o
r
m
an
ce
f
o
r
lap
to
p
s
)
,
m
ak
in
g
it
s
u
itab
le
f
o
r
f
in
e
-
g
r
ain
ed
s
en
tim
en
t
class
if
icatio
n
at
th
e
asp
ec
t
lev
el.
I
t
is
co
m
m
o
n
l
y
u
s
ed
f
o
r
tr
ai
n
in
g
an
d
ev
alu
at
in
g
AB
S
A
m
o
d
els.
T
h
e
Sen
tSem
Sy
n
Net
m
o
d
el
was
im
p
lan
ted
o
n
a
s
y
s
tem
h
av
in
g
W
in
d
o
w
1
1
o
p
er
atin
g
s
y
s
tem
,
with
1
6
GM
R
AM
.
T
h
e
s
y
s
tem
co
n
s
id
er
ed
h
ad
AM
D
R
y
ze
n
5
Pro
ce
s
s
o
r
an
d
was
r
u
n
n
in
g
with
g
r
ap
h
i
ca
l
p
r
o
ce
s
s
in
g
u
n
it
,
i.e
.
,
4
GB
NVI
DA
GeFo
r
ce
GT
X1
6
5
0
f
o
r
f
aster
p
r
o
ce
s
s
in
g
.
Fo
r
ex
ec
u
tio
n
,
th
e
Py
th
o
n
en
v
i
r
o
n
m
e
n
t
was
u
s
ed
.
Fo
r
ev
al
u
atio
n
,
two
s
tan
d
ar
d
m
etr
ics
u
s
ed
f
o
r
AB
S
A
ev
alu
atio
n
wer
e
co
n
s
id
er
e
d
,
i.e
.
,
ac
c
u
r
ac
y
a
n
d
m
ac
r
o
-
F
-
s
co
r
e.
T
h
e
ac
cu
r
ac
y
an
d
m
ac
r
o
-
F
-
s
co
r
e
is
ev
alu
ate
d
u
s
in
g
(
1
7
)
a
n
d
(
1
8
)
.
=
+
+
+
+
(
1
7
)
−
−
=
∑
1
−
=
1
(
1
8
)
I
n
(
1
7
)
,
d
en
o
tes
tr
u
e
p
o
s
itiv
e,
d
en
o
tes
tr
u
e
n
e
g
ativ
e,
d
en
o
tes
f
alse
p
o
s
itiv
e
,
an
d
d
en
o
tes
f
alse
n
eg
ativ
e.
I
n
(
1
8
)
,
F1
-
s
co
r
e
is
ev
alu
ated
u
s
in
g
(
1
9
)
.
I
n
(
1
9
)
,
an
d
ar
e
ev
alu
ate
d
u
s
in
g
(
2
0
)
an
d
(
2
1
)
.
1
−
=
2
×
×
+
(
1
9
)
=
+
(
2
0
)
=
+
(
2
1
)
4
.
1
.
SentSem
Sy
nNet
wo
r
k
perf
o
rma
nce
o
n Sem
E
v
a
l2
0
1
4
T
a
s
k
4
da
t
a
s
et
T
h
e
ev
al
u
atio
n
r
esu
lts
o
f
Sen
tSem
Sy
n
Net
m
o
d
el
o
n
Sem
E
v
al
2
0
1
4
T
ask
4
d
ataset
is
p
r
esen
ted
in
Fig
u
r
e
2
.
T
h
e
f
i
n
d
in
g
s
s
h
o
w
ef
f
ec
tiv
en
ess
in
AB
SA
f
o
r
b
o
th
lap
to
p
an
d
r
estau
r
a
n
t
d
o
m
ain
s
.
Fo
r
r
estau
r
a
n
t
d
ataset,
Sen
tSem
Sy
n
Net
h
as
ac
h
iev
ed
8
8
.
2
5
%
ac
cu
r
ac
y
an
d
8
2
.
9
5
%
m
ac
r
o
-
F
-
s
co
r
e.
Fo
r
lap
to
p
d
ataset,
Sen
tSem
Sy
n
Net
ac
h
iev
ed
8
4
.
5
2
%
ac
cu
r
ac
y
a
n
d
8
0
.
2
6
m
ac
r
o
-
F
-
s
co
r
e.
T
h
e
f
in
d
in
g
s
s
h
o
w
th
at
th
e
in
teg
r
atio
n
o
f
b
o
t
h
s
em
an
tic
an
d
s
y
n
tactic
in
f
o
r
m
atio
n
u
s
in
g
g
r
ap
h
-
b
ased
DL
a
p
p
r
o
ac
h
i
m
p
r
o
v
es
s
en
tim
en
t
class
if
icatio
n
.
Al
s
o
,
th
e
f
in
d
in
g
s
s
h
o
w
th
at
th
e
Sen
tSem
S
y
n
Net
ef
f
ec
tiv
ely
ca
p
tu
r
es
im
p
o
r
tan
t
r
elatio
n
s
h
i
p
b
etwe
en
co
n
tex
t
a
n
d
asp
ec
t
t
er
m
s
.
Fu
r
th
er
,
in
th
e
n
ex
t
s
ec
tio
n
,
th
e
Sen
Sem
Sy
n
Net
ap
p
r
o
ac
h
is
co
m
p
ar
ed
with
ex
is
tin
g
ap
p
r
o
ac
h
es.
Fig
u
r
e
2
.
Sen
tSem
Sy
n
Net
p
e
r
f
o
r
m
an
ce
o
n
Sem
E
v
al
T
ask
4
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
8
1
4
-
1
8
2
4
1822
4
.
2
.
Co
m
pa
ra
t
iv
e
s
t
ud
y
T
h
e
co
m
p
ar
ativ
e
r
esu
lts
co
m
p
ar
in
g
th
e
ex
is
tin
g
AB
SA
ap
p
r
o
ac
h
es
p
r
esen
ted
in
th
e
liter
atu
r
e
s
u
r
v
e
y
h
as
b
ee
n
co
m
p
ar
ed
with
p
r
o
p
o
s
ed
Sen
Sem
Sy
n
Net
as
p
r
esen
ted
in
T
ab
le
2
.
Fo
r
r
estau
r
an
t
d
ataset,
Sen
tSem
Sy
n
Net
ac
h
iev
ed
8
8
.
2
5
%
ac
cu
r
ac
y
an
d
8
2
.
9
5
%
m
ac
r
o
-
F
-
s
co
r
e,
o
u
tp
er
f
o
r
m
i
n
g
SAGC
N
wh
ich
ac
h
iev
ed
8
7
.
5
3
%
ac
cu
r
ac
y
an
d
8
1
.
2
8
%
m
ac
r
o
-
F
-
s
co
r
e
an
d
B
I
C
-
GC
N
wh
ich
ac
h
iev
ed
8
7
.
1
6
%
ac
cu
r
ac
y
an
d
8
1
.
2
4
%
m
ac
r
o
-
F
-
s
co
r
e.
T
h
is
i
s
b
ec
au
s
e
Sen
Sem
Sy
n
Net
ap
p
r
o
ac
h
c
o
n
s
id
er
ed
jo
in
tly
in
teg
r
atin
g
s
em
an
tic
a
n
d
s
y
n
tactic
f
ea
tu
r
es
u
s
in
g
Se
n
tSem
Sy
n
Net
g
r
ap
h
-
b
ased
s
tr
u
ctu
r
e,
wh
ic
h
allo
wed
f
o
r
ac
h
iev
in
g
b
etter
r
ep
r
esen
tatio
n
o
f
c
o
n
tex
tu
al
d
ep
en
d
e
n
cy
an
d
s
en
tim
en
t
-
r
e
lated
in
f
o
r
m
atio
n
.
Fu
r
th
er
,
f
o
r
lap
to
p
d
ataset,
Sen
tSem
Sy
n
Net
ac
h
iev
ed
8
4
.
5
2
%
ac
cu
r
ac
y
a
n
d
8
0
.
2
6
%
m
ac
r
o
-
F
-
s
co
r
e,
wh
er
ea
s
SAGC
N
ac
h
iev
ed
8
3
.
0
6
%
ac
cu
r
ac
y
a
n
d
7
9
.
6
9
%
ac
cu
r
ac
y
an
d
B
I
C
-
GC
N
ac
h
iev
ed
8
1
.
7
0
%
ac
cu
r
ac
y
an
d
7
8
.
4
5
%
m
ac
r
o
-
F
-
s
co
r
e.
T
h
e
o
th
er
ap
p
r
o
ac
h
es
lik
ed
MO
L
-
AB
S
A
an
d
L
GC
F
ac
h
iev
ed
b
etter
r
esu
lts
,
b
u
t
f
ailed
to
c
o
n
s
id
er
in
teg
r
atio
n
o
f
s
y
n
tactic
an
d
s
em
an
tic
in
f
o
r
m
atio
n
at
g
r
a
p
h
-
le
v
el,
wh
ich
lim
ited
th
eir
class
i
f
icatio
n
ac
cu
r
ac
y
.
Sen
tSem
Sy
n
Net’
s
ab
ilit
y
to
m
ain
tain
m
ac
r
o
-
F
-
s
co
r
e
al
o
n
g
s
id
e
h
ig
h
ac
cu
r
ac
y
c
o
n
f
i
r
m
s
its
s
tr
en
g
th
i
n
h
an
d
lin
g
class
im
b
alan
ce
an
d
s
u
b
tle
s
en
tim
en
t
i
m
p
o
r
tan
ce
.
T
h
e
r
esu
lts
cl
ea
r
ly
in
d
icate
th
at
Sen
tSem
Sy
n
Net
p
r
o
v
id
es
b
ett
er
AB
S
A
f
r
am
ewo
r
k
b
y
ef
f
ec
tiv
ely
lev
er
ag
in
g
b
o
th
d
e
p
en
d
en
cy
r
elatio
n
s
an
d
s
em
an
tic
s
im
ilar
ity
.
T
ab
le
2
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
Sen
tSem
Sy
n
Net
with
e
x
is
tin
g
AB
SA m
o
d
els
Ref
M
o
d
e
l
S
e
mEv
a
l
2
0
1
4
r
e
st
a
u
r
a
n
t
S
e
mEv
a
l
2
0
1
4
l
a
p
t
o
p
A
c
c
u
r
a
c
y
M
a
c
r
o
F
-
s
c
o
r
e
A
c
c
u
r
a
c
y
M
a
c
r
o
F
-
s
c
o
r
e
[
1
6
]
C
P
A
-
SA
8
2
.
6
4
7
3
.
3
8
7
5
.
1
8
7
1
.
5
[
1
7
]
H
e
t
e
-
GNN
8
1
.
9
1
7
3
.
7
4
7
4
.
0
8
6
9
.
4
5
[
1
8
]
LG
C
F
-
C
D
M
-
C
D
W
8
5
.
5
2
7
9
.
8
5
8
1
.
2
9
7
8
.
8
6
[
2
0
]
S
D
TG
C
N
8
3
.
8
2
7
6
.
1
3
7
8
.
6
4
7
5
.
5
[
2
1
]
S
A
G
C
N
8
7
.
5
3
8
1
.
2
8
8
3
.
0
6
7
9
.
6
9
[
2
2
]
K
H
G
C
N
-
-
8
0
.
8
7
7
7
.
9
[
2
3
]
SS
-
G
C
N
8
2
.
9
6
7
4
.
2
6
7
5
.
8
6
7
1
.
7
8
[
2
4
]
D
C
A
S
A
M
8
6
.
7
8
1
.
1
9
8
0
.
5
6
77
[
2
5
]
M
O
L
-
A
B
S
A
8
5
.
2
6
7
9
.
2
7
8
1
.
1
9
7
8
.
5
2
[
2
6
]
B
I
C
-
G
C
N
8
7
.
1
6
8
1
.
2
4
8
1
.
7
7
8
.
4
5
[
2
7
]
SM
-
G
C
N
+
B
ER
T
8
6
.
3
1
7
9
.
2
8
0
.
8
6
7
6
.
7
S
e
n
t
S
e
mS
y
n
N
e
t
8
8
.
2
5
8
2
.
9
5
8
4
.
5
2
8
0
.
2
6
5.
CO
NCLU
SI
O
N
I
n
th
is
w
o
r
k
,
a
n
o
v
el
m
o
d
el
ca
lled
Sen
tSem
Sy
n
Net
was
p
r
o
p
o
s
ed
to
en
h
a
n
ce
th
e
p
e
r
f
o
r
m
an
ce
o
f
AB
SA,
a
s
u
b
d
o
m
ain
o
f
s
en
ti
m
en
t
an
aly
s
is
th
at
id
en
tifie
s
s
en
tim
en
ts
ex
p
r
ess
ed
to
war
d
s
s
p
ec
if
ic
asp
ec
ts
in
tex
t.
T
r
ad
itio
n
al
m
o
d
els
o
f
te
n
f
ail
to
ef
f
ec
tiv
el
y
in
teg
r
ate
s
em
an
tic
an
d
s
y
n
tactic
in
f
o
r
m
atio
n
in
a
u
n
if
ied
m
an
n
er
,
wh
ich
lim
its
th
eir
ca
p
ac
ity
to
co
r
r
ec
tly
ca
p
t
u
r
e
s
en
ti
m
en
t im
p
o
r
tan
c
e,
esp
ec
ially
in
co
m
p
lex
s
en
ten
ce
s
tr
u
ctu
r
es.
T
o
o
v
er
co
m
e
th
is
,
th
is
wo
r
k
p
r
esen
ted
a
g
r
ap
h
-
b
ased
DL
m
o
d
el
th
at
s
im
u
ltan
eo
u
s
ly
u
tili
ze
s
s
y
n
tactic
d
ep
en
d
en
cies
an
d
s
em
an
tic
ass
o
ciatio
n
s
b
y
co
n
s
tr
u
ctin
g
a
h
y
b
r
id
g
r
ap
h
r
ep
r
ese
n
tatio
n
o
f
s
en
ten
ce
s
.
T
h
e
Sen
tSem
Sy
n
Net
m
o
d
el
was
ev
alu
ated
o
n
t
h
e
Sem
E
v
al
2
0
1
4
T
ask
4
d
atasets
f
o
r
r
estau
r
an
t
an
d
la
p
to
p
d
o
m
ain
s
.
I
t
ac
h
iev
ed
8
8
.
2
5
%
ac
cu
r
ac
y
an
d
8
2
.
9
5
%
m
ac
r
o
-
F
-
s
co
r
e
o
n
r
estau
r
a
n
t
d
ataset,
an
d
8
4
.
5
2
%
ac
cu
r
ac
y
a
n
d
8
0
.
2
6
%
m
ac
r
o
-
F
-
s
co
r
e
o
f
o
n
lp
to
p
d
ataset.
T
h
ese
r
esu
lts
s
u
r
p
ass
ex
is
tin
g
m
o
d
els
s
u
ch
as
SAGC
N,
B
I
C
-
G
C
N,
an
d
L
GC
F
,
d
em
o
n
s
tr
atin
g
th
at
th
e
f
u
s
io
n
o
f
s
em
an
tic
an
d
s
y
n
tactic
f
ea
tu
r
es si
g
n
if
ican
tly
en
h
an
ce
s
s
en
tim
en
t
class
if
icatio
n
ac
cu
r
ac
y
a
n
d
r
o
b
u
s
tn
e
s
s
ac
r
o
s
s
d
atase
ts
.
I
n
co
n
clu
s
io
n
,
th
e
p
r
o
p
o
s
ed
Sen
tSem
Sy
n
Net
m
o
d
el
ef
f
ec
tiv
ely
ad
d
r
ess
es
th
e
lim
itatio
n
s
o
f
p
r
io
r
AB
SA
m
o
d
els
b
y
in
teg
r
atin
g
d
ee
p
s
y
n
tactic
s
tr
u
ctu
r
es
with
s
em
an
tic
m
ea
n
in
g
,
lead
in
g
to
im
p
r
o
v
ed
asp
ec
t
-
le
v
el
s
en
tim
en
t
cl
ass
if
icatio
n
.
Fu
tu
r
e
wo
r
k
will
f
o
cu
s
o
n
in
c
o
r
p
o
r
ati
n
g
s
ar
ca
s
m
id
en
tific
atio
n
,
a
c
h
allen
g
in
g
b
u
t
cr
itical
f
ea
t
u
r
e
th
at
o
f
ten
alter
s
th
e
tr
u
e
s
en
tim
en
t o
f
a
s
tatem
en
t,
f
u
r
th
er
e
n
h
an
ci
n
g
th
e
m
o
d
el’
s
r
ea
l
-
wo
r
ld
ap
p
licab
ilit
y
.
ACK
NO
WL
E
DG
E
M
E
NT
Au
th
o
r
s
wo
u
ld
lik
e
to
ex
p
r
e
s
s
o
u
r
h
ea
r
tf
elt
th
an
k
s
to
o
u
r
g
u
id
e
f
o
r
h
im
u
n
wav
er
in
g
g
u
id
an
ce
,
in
v
alu
ab
le
in
s
ig
h
ts
,
an
d
en
c
o
u
r
ag
em
en
t th
r
o
u
g
h
o
u
t th
e
r
esea
r
ch
p
r
o
ce
s
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
No
f
u
n
d
in
g
is
r
aised
f
o
r
th
is
r
e
s
ea
r
ch
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
ema
n
tic
-
s
yn
ta
ctic
g
r
a
p
h
n
et
w
o
r
k
fo
r
a
s
p
ec
t
-
b
a
s
ed
s
en
timen
t a
n
a
lysi
s
(
R
ek
h
a
B
d
u
r
g
a
Ha
r
is
h
)
1823
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
t
r
ib
u
to
r
R
o
les
T
a
x
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
i
d
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
R
ek
h
a
B
d
u
r
g
a
Har
is
h
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
N
e
e
l
a
m
b
i
k
e
S
i
d
d
a
l
i
n
g
a
i
a
h
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
T
h
e
au
th
o
r
d
ec
lar
es n
o
co
n
f
lict o
f
in
ter
est
.
DATA AV
AI
L
AB
I
L
I
T
Y
Data
s
ets u
tili
ze
d
in
th
is
r
esear
ch
ar
e
cited
in
r
ef
er
e
n
ce
[
1
7
]
.
RE
F
E
R
E
NC
E
S
[
1
]
K
.
P
a
r
k
,
S
.
P
a
r
k
,
a
n
d
J
.
Jo
u
n
g
,
“
C
o
n
t
e
x
t
u
a
l
m
e
a
n
i
n
g
-
b
a
se
d
a
p
p
r
o
a
c
h
t
o
f
i
n
e
-
g
r
a
i
n
e
d
o
n
l
i
n
e
p
r
o
d
u
c
t
r
e
v
i
e
w
a
n
a
l
y
s
i
s
f
o
r
p
r
o
d
u
c
t
d
e
s
i
g
n
,
”
I
EE
E
A
c
c
e
ss
,
v
o
l
.
1
2
,
p
p
.
4
2
2
5
–
4
2
3
8
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ES
S
.
2
0
2
3
.
3
3
4
3
5
0
1
.
[
2
]
D
.
A
man
g
e
l
d
i
,
A
.
U
sm
a
n
o
v
a
,
a
n
d
P
.
S
h
a
m
o
i
,
“
U
n
d
e
r
st
a
n
d
i
n
g
e
n
v
i
r
o
n
m
e
n
t
a
l
p
o
st
s
:
se
n
t
i
me
n
t
a
n
d
e
m
o
t
i
o
n
a
n
a
l
y
si
s
o
f
so
c
i
a
l
med
i
a
d
a
t
a
,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
1
2
,
p
p
.
3
3
5
0
4
–
3
3
5
2
3
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
4
.
3
3
7
1
5
8
5
.
[
3
]
V
.
A
.
A
.
Q
u
i
r
i
t
a
,
J
.
D
.
C
á
r
d
e
n
a
s,
W
.
A
l
i
a
g
a
,
V
.
H
.
A
.
Q
u
i
r
i
t
a
,
A
.
P
a
l
a
c
i
o
s,
a
n
d
R
.
B
.
S
i
e
r
r
a
,
“
P
e
r
u
v
i
a
n
p
r
e
si
d
e
n
t
i
a
l
d
e
b
a
t
e
s
i
n
t
h
e
e
l
e
c
t
i
o
n
s
o
f
2
0
2
1
i
n
Tw
i
t
t
e
r
/
X
:
a
se
n
t
i
me
n
t
a
n
a
l
y
si
s
a
p
p
r
o
a
c
h
,
”
I
EEE
A
c
c
e
ss
,
v
o
l
.
1
2
,
p
p
.
1
3
8
3
8
6
–
1
3
8
3
9
8
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
4
.
3
4
6
5
5
0
3
.
[
4
]
S
.
S
h
a
n
,
J.
S
u
n
,
a
n
d
R
.
M
.
C
.
M
a
c
a
w
i
l
e
,
“
Ex
a
mi
n
i
n
g
c
u
s
t
o
mer
sa
t
i
sf
a
c
t
i
o
n
t
h
r
o
u
g
h
t
r
a
n
sf
o
r
mer
-
b
a
s
e
d
s
e
n
t
i
men
t
a
n
a
l
y
si
s
f
o
r
i
mp
r
o
v
i
n
g
b
i
l
i
n
g
u
a
l
e
-
c
o
mm
e
r
c
e
e
x
p
e
r
i
e
n
c
e
s,”
I
EE
E
A
c
c
e
ss
,
v
o
l
.
1
3
,
p
p
.
5
1
0
2
7
–
5
1
0
4
1
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
5
1
6
6
6
.
[
5
]
W
.
A
h
m
a
d
,
H
.
U
.
K
h
a
n
,
F
.
K
.
A
l
a
r
f
a
j
,
a
n
d
M
.
A
l
r
e
s
h
o
o
d
i
,
“
A
sp
e
c
t
-
b
a
se
d
sen
t
i
m
e
n
t
a
n
a
l
y
si
s
:
a
c
o
m
p
r
e
h
e
n
si
v
e
r
e
v
i
e
w
a
n
d
o
p
e
n
r
e
sea
r
c
h
c
h
a
l
l
e
n
g
e
s,
”
I
EE
E
A
c
c
e
s
s
,
v
o
l
.
1
3
,
p
p
.
6
5
1
3
8
–
6
5
1
8
2
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
5
5
7
4
4
.
[
6
]
D
.
S
i
n
g
h
,
S
.
S
.
B
a
r
v
e
,
a
n
d
A
.
K
.
D
w
i
v
e
d
i
,
“
O
p
t
i
A
S
A
R
:
o
p
t
i
m
i
z
e
d
a
sp
e
c
t
-
b
a
se
d
s
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
o
f
r
e
v
i
e
w
s
w
i
t
h
B
i
LST
M
-
G
R
U
a
n
d
N
E
R
-
B
E
R
T
i
n
h
e
a
l
t
h
c
a
r
e
d
e
c
i
si
o
n
-
ma
k
i
n
g
,
”
I
E
EE
Ac
c
e
s
s
,
v
o
l
.
1
3
,
p
p
.
4
7
4
5
9
–
4
7
4
7
4
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
4
9
3
0
3
.
[
7
]
P
.
Z
h
a
n
g
,
M
.
F
u
,
R
.
Z
h
a
o
,
H
.
Z
h
a
n
g
,
a
n
d
C
.
L
u
o
,
“
P
U
R
E
:
p
e
r
s
o
n
a
l
i
t
y
-
c
o
u
p
l
e
d
m
u
l
t
i
-
t
a
s
k
l
e
a
r
n
i
n
g
f
r
a
m
e
w
o
r
k
f
o
r
a
s
p
e
c
t
-
b
a
se
d
mu
l
t
i
m
o
d
a
l
s
e
n
t
i
me
n
t
a
n
a
l
y
si
s,
”
I
EE
E
T
r
a
n
s
a
c
t
i
o
n
s
o
n
K
n
o
w
l
e
d
g
e
a
n
d
D
a
t
a
En
g
i
n
e
e
r
i
n
g
,
v
o
l
.
3
7
,
n
o
.
1
,
p
p
.
4
6
2
–
4
7
7
,
J
a
n
.
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
T
K
D
E.
2
0
2
4
.
3
4
8
5
1
0
8
.
[
8
]
R
.
L
i
,
H
.
C
h
e
n
,
F
.
F
e
n
g
,
Z.
M
a
,
X
.
W
a
n
g
,
a
n
d
E.
H
o
v
y
,
“
D
u
a
l
G
C
N
:
e
x
p
l
o
r
i
n
g
sy
n
t
a
c
t
i
c
a
n
d
s
e
ma
n
t
i
c
i
n
f
o
r
mat
i
o
n
f
o
r
a
s
p
e
c
t
-
b
a
s
e
d
se
n
t
i
me
n
t
a
n
a
l
y
si
s
,
”
I
EEE
T
ra
n
sa
c
t
i
o
n
s
o
n
N
e
u
r
a
l
N
e
t
w
o
rks
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
5
,
n
o
.
6
,
p
p
.
7
6
4
2
–
7
6
5
6
,
J
u
n
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS.
2
0
2
2
.
3
2
1
9
6
1
5
.
[
9
]
M
.
G
u
a
n
,
F
.
Li
,
a
n
d
Y
.
X
u
e
,
“
En
h
a
n
c
e
d
sy
n
t
a
c
t
i
c
a
n
d
s
e
ma
n
t
i
c
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
w
i
t
h
c
o
n
t
r
a
st
i
v
e
l
e
a
r
n
i
n
g
f
o
r
a
s
p
e
c
t
-
b
a
s
e
d
se
n
t
i
me
n
t
a
n
a
l
y
s
i
s,”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
S
o
c
i
a
l
S
y
s
t
e
m
s
,
v
o
l
.
1
1
,
n
o
.
1
,
p
p
.
8
5
9
–
8
7
0
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TC
S
S
.
2
0
2
3
.
3
2
9
6
4
7
6
.
[
1
0
]
A
.
T.
H
a
r
y
o
n
o
,
R
.
S
a
r
n
o
,
R
.
N
.
E.
A
n
g
g
r
a
i
n
i
,
a
n
d
K
.
R
.
S
u
n
g
k
o
n
o
,
“
P
e
r
mu
t
e
d
t
e
m
p
o
r
a
l
k
o
l
m
o
g
o
r
o
v
-
a
r
n
o
l
d
n
e
t
w
o
r
k
s
f
o
r
s
t
o
c
k
p
r
i
c
e
f
o
r
e
c
a
s
t
i
n
g
u
si
n
g
g
e
n
e
r
a
t
i
v
e
a
sp
e
c
t
-
b
a
s
e
d
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
1
2
,
p
p
.
1
7
8
6
7
2
–
1
7
8
6
8
9
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
4
.
3
5
0
6
6
5
8
.
[
1
1
]
S
.
G
u
p
t
a
e
t
a
l
.
,
“
A
sp
e
c
t
b
a
se
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
i
n
s
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
u
si
n
g
B
i
-
G
R
U
-
LST
M
m
o
d
e
l
,
”
J
o
u
rn
a
l
o
f
M
o
b
i
l
e
Mu
l
t
i
m
e
d
i
a
,
v
o
l
.
2
0
,
n
o
.
4
,
p
p
.
9
3
5
–
9
6
0
,
O
c
t
.
2
0
2
4
,
d
o
i
:
1
0
.
1
3
0
5
2
/
j
m
m1
5
5
0
-
4
6
4
6
.
2
0
4
8
.
[
1
2
]
W
.
S
h
u
,
C
.
Z
h
a
i
,
a
n
d
K
.
Y
u
,
“
D
A
G
C
N
:
d
u
a
l
-
c
h
a
n
n
e
l
a
n
d
a
sp
e
c
t
-
a
w
a
r
e
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
f
o
r
a
s
p
e
c
t
-
b
a
s
e
d
s
e
n
t
i
me
n
t
a
n
a
l
y
si
s
i
n
c
o
mp
u
t
a
t
i
o
n
a
l
s
o
c
i
a
l
sy
s
t
e
ms,”
I
E
EE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
S
o
c
i
a
l
S
y
st
e
m
s
,
v
o
l
.
1
2
,
n
o
.
6
,
p
p
.
5
4
0
5
–
5
4
1
5
,
D
e
c
.
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
T
C
S
S
.
2
0
2
4
.
3
4
1
8
4
7
2
.
[
1
3
]
S
.
S
.
M
a
l
i
k
e
t
a
l
.
,
“
M
u
l
t
i
-
mo
d
a
l
e
m
o
t
i
o
n
d
e
t
e
c
t
i
o
n
a
n
d
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s,”
I
E
EE
Ac
c
e
ss
,
v
o
l
.
1
3
,
p
p
.
5
9
7
9
0
–
5
9
8
1
0
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
5
2
4
7
5
.
[
1
4
]
H
.
T.
P
h
a
n
,
V
.
D
.
N
g
u
y
e
n
,
a
n
d
N
.
T.
N
g
u
y
e
n
,
“
M
u
l
G
C
N
:
m
u
l
t
i
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
f
o
r
a
sp
e
c
t
-
l
e
v
e
l
s
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
,
”
I
EEE
A
c
c
e
ss
,
v
o
l
.
1
3
,
p
p
.
2
6
3
0
4
–
2
6
3
1
7
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
3
7
3
4
0
.
[
1
5
]
C
.
W
u
a
n
d
H
.
K
a
n
g
,
“
H
a
r
n
e
ss
i
n
g
s
y
n
t
a
x
g
c
n
a
n
d
m
u
l
t
i
-
v
i
e
w
i
n
t
e
r
a
c
t
i
o
n
f
o
r
c
o
n
v
e
r
sat
i
o
n
a
l
a
s
p
e
c
t
-
b
a
se
d
q
u
a
d
r
u
p
l
e
se
n
t
i
me
n
t
a
n
a
l
y
si
s
,
”
I
E
EE
Ac
c
e
ss
,
v
o
l
.
1
3
,
p
p
.
7
3
3
3
2
–
7
3
3
4
1
,
2
0
2
5
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
5
.
3
5
6
2
9
6
7
.
[
1
6
]
B
.
H
u
a
n
g
e
t
a
l
.
,
“
A
s
p
e
c
t
-
l
e
v
e
l
se
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
w
i
t
h
a
s
p
e
c
t
-
sp
e
c
i
f
i
c
c
o
n
t
e
x
t
p
o
si
t
i
o
n
i
n
f
o
r
m
a
t
i
o
n
,
”
K
n
o
w
l
e
d
g
e
-
Ba
se
d
S
y
s
t
e
m
s
,
v
o
l
.
2
4
3
,
M
a
y
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
k
n
o
sy
s
.
2
0
2
2
.
1
0
8
4
7
3
.
[
1
7
]
G
.
Lu
,
J.
L
i
,
a
n
d
J.
W
e
i
,
“
A
sp
e
c
t
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
w
i
t
h
h
e
t
e
r
o
g
e
n
e
o
u
s
g
r
a
p
h
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
I
n
f
o
rm
a
t
i
o
n
Pro
c
e
ss
i
n
g
&
Ma
n
a
g
e
m
e
n
t
,
v
o
l
.
5
9
,
n
o
.
4
,
J
u
l
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
p
m.
2
0
2
2
.
1
0
2
9
5
3
.
[
1
8
]
T.
L
i
n
a
n
d
I
.
Jo
e
,
“
A
n
a
d
a
p
t
i
v
e
m
a
s
k
e
d
a
t
t
e
n
t
i
o
n
m
e
c
h
a
n
i
s
m
t
o
a
c
t
o
n
t
h
e
l
o
c
a
l
t
e
x
t
i
n
a
g
l
o
b
a
l
c
o
n
t
e
x
t
f
o
r
a
sp
e
c
t
-
b
a
s
e
d
se
n
t
i
me
n
t
a
n
a
l
y
si
s
,
”
I
E
EE
Ac
c
e
ss
,
v
o
l
.
1
1
,
p
p
.
4
3
0
5
5
–
4
3
0
6
6
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
3
.
3
2
7
0
9
2
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.