I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
41
,
No
.
3
,
Ma
r
ch
20
2
6
,
p
p
.
95
4
~
96
5
I
SS
N:
2
502
-
4
7
52
,
DOI
: 1
0
.
1
1
5
9
1
/ijee
cs
.v
41
.
i
3
.
pp
954
-
9
6
5
954
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs
.
ia
esco
r
e.
co
m
A hybrid a
pp
ro
a
ch f
o
r meas
uring
s
ema
ntic
sim
ila
rit
y
in
lex
ica
lly
iden
tical
but
a
mbig
uo
us se
ntences
B
t
is
s
a
m
E
l J
a
na
t
i
1
,
Adil E
na
a
na
i
2
,
F
a
do
ua
G
ha
nim
i
1
1
D
e
p
a
r
t
me
n
t
o
f
P
h
y
si
c
s E
n
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
S
c
i
e
n
c
e
,
I
b
n
T
o
f
a
i
l
U
n
i
v
e
r
si
t
y
,
K
e
n
i
t
r
a
,
M
o
r
o
c
c
o
2
D
e
p
a
r
t
me
n
t
o
f
i
n
f
o
r
m
a
t
i
c
,
A
b
d
e
l
ma
l
e
k
S
a
a
d
i
U
n
i
v
e
r
si
t
y
,
Té
t
o
u
a
n
,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
1
9
,
2
0
2
5
R
ev
is
ed
Dec
5
,
2
0
2
5
Acc
ep
ted
Dec
1
3
,
2
0
2
5
Th
is
stu
d
y
a
d
d
re
ss
e
s
th
e
c
rit
ica
l
c
h
a
ll
e
n
g
e
o
f
se
m
a
n
ti
c
sim
il
a
rit
y
a
n
d
lex
ica
l
d
isa
m
b
ig
u
a
t
io
n
in
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
,
fo
c
u
sin
g
o
n
se
n
te
n
c
e
s
with
stru
c
tu
ra
l
a
n
d
lex
ica
l
a
m
b
ig
u
it
ies
.
We
i
n
tr
o
d
u
c
e
a
n
in
n
o
v
a
ti
v
e
h
y
b
r
id
a
p
p
ro
a
c
h
t
h
a
t
sy
n
e
rg
isti
c
a
ll
y
c
o
m
b
in
e
s
sy
m
b
o
li
c
a
n
d
n
e
u
ra
l
m
e
th
o
d
s
t
o
b
e
tt
e
r
a
li
g
n
with
h
u
m
a
n
j
u
d
g
m
e
n
t.
Ou
r
m
e
th
o
d
o
lo
g
y
d
y
n
a
m
ica
ll
y
in
teg
ra
tes
fu
z
z
y
Ja
c
c
a
rd
’
s
lex
ica
l
p
re
c
isi
o
n
wit
h
S
BERT
e
m
b
e
d
d
in
g
s
’
c
o
n
tex
t
u
a
l
se
n
siti
v
it
y
,
e
n
a
b
li
n
g
a
d
a
p
ti
v
e
se
m
a
n
ti
c
a
m
b
ig
u
i
ty
re
so
lu
ti
o
n
.
E
x
p
e
rime
n
tal
e
v
a
lu
a
ti
o
n
o
n
3
3
a
m
b
ig
u
o
u
s
s
e
n
ten
c
e
s
d
e
m
o
n
stra
tes
th
a
t
o
u
r
a
p
p
r
o
a
c
h
sig
n
ifi
c
a
n
t
ly
o
u
t
p
e
rfo
rm
s
c
o
n
v
e
n
ti
o
n
a
l
a
rti
ficia
l
in
telli
g
e
n
c
e
(AI)
sy
ste
m
s,
a
c
h
iev
in
g
a
n
1
1
.
7
%
re
d
u
c
ti
o
n
in
m
e
a
n
a
b
so
lu
te
e
rro
r
c
o
m
p
a
re
d
t
o
re
fe
re
n
c
e
m
o
d
e
ls,
wit
h
sta
ti
stica
l
a
n
a
ly
sis
c
o
n
firmin
g
r
o
b
u
st
re
su
lt
s
(d
=
-
0
.
8
0
,
p
<
0
.
0
0
1
)
.
T
h
is
re
p
re
se
n
ts
a
6
5
%
im
p
ro
v
e
m
e
n
t
in
h
u
m
a
n
e
v
a
l
u
a
ti
o
n
a
li
g
n
m
e
n
t
o
v
e
r
e
x
isti
n
g
m
e
th
o
d
s.
O
u
r
re
se
a
rc
h
c
o
n
tri
b
u
tes
t
o
a
d
v
a
n
c
i
n
g
t
h
e
field
b
y
sh
o
win
g
t
h
a
t
a
rc
h
i
tec
tu
ra
l
i
n
telli
g
e
n
c
e
c
a
n
su
r
p
a
ss
m
e
re
p
a
ra
m
e
ter
sc
a
li
n
g
,
o
ffe
rin
g
a
n
e
ffe
c
ti
v
e
so
l
u
ti
o
n
f
o
r
a
p
p
li
c
a
ti
o
n
s
re
q
u
iri
n
g
b
o
t
h
p
re
c
isio
n
a
n
d
in
terp
re
tab
il
it
y
,
wit
h
p
ro
m
isi
n
g
d
irec
ti
o
n
s
fo
r
m
u
lt
il
in
g
u
a
l
e
x
te
n
sio
n
a
n
d
e
x
p
lain
a
b
le AI
in
te
g
ra
ti
o
n
.
K
ey
w
o
r
d
s
:
Fre
n
ch
NL
P
L
in
g
u
is
tic
f
ea
tu
r
es
Sen
ten
ce
s
im
ilar
ity
NL
P
T
ex
tu
al
s
em
an
tic
s
im
ilar
ity
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
:
B
tis
s
am
E
l Jan
ati
Dep
ar
tm
en
t o
f
P
h
y
s
ics,
Facu
lty
o
f
Scien
ce
,
I
b
n
T
o
f
ail
Un
iv
e
r
s
ity
K
en
itra
,
Mo
r
r
o
c
o
E
m
ail: Bt
is
s
am
.
eljan
ati@
u
it.a
c.
m
a
1.
I
NT
RO
D
UCT
I
O
N
Natu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
NL
P)
en
ab
les
m
ac
h
in
es
to
i
n
ter
p
r
et
a
n
d
g
en
e
r
ate
h
u
m
an
lan
g
u
a
g
e,
f
o
r
m
in
g
th
e
f
o
u
n
d
atio
n
o
f
ap
p
licatio
n
s
s
u
ch
as
s
ea
r
ch
en
g
in
es,
m
ac
h
in
e
tr
an
s
latio
n
,
a
n
d
d
ialo
g
u
e
s
y
s
tem
s
.
Am
o
n
g
its
co
r
e
ch
allen
g
es,
s
em
an
tic
tex
tu
al
s
im
ilar
ity
(
S
T
S)
aim
s
to
q
u
an
tify
h
o
w
clo
s
ely
two
s
en
ten
ce
s
co
n
v
ey
th
e
s
am
e
m
ea
n
in
g
.
W
h
ile
s
ig
n
if
ican
t
p
r
o
g
r
ess
h
as
b
ee
n
m
ad
e
in
E
n
g
lis
h
N
L
P,
Fre
n
ch
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
f
ac
es
u
n
iq
u
e
c
h
a
llen
g
es
d
u
e
to
its
r
ich
m
o
r
p
h
o
lo
g
y
,
g
r
a
m
m
atica
l
co
m
p
l
ex
ity
,
an
d
lim
ited
an
n
o
tated
r
eso
u
r
ce
s
.
E
x
is
tin
g
ap
p
r
o
ac
h
es
f
o
r
Fre
n
ch
s
em
an
tic
s
im
ilar
ity
r
ev
ea
l
cr
itical
lim
itatio
n
s
.
Neu
r
al
m
o
d
els
lik
e
C
am
em
B
E
R
T
an
d
Flau
B
E
R
T
,
wh
ile
ef
f
ec
tiv
e
f
o
r
g
en
er
al
task
s
,
o
f
ten
s
tr
u
g
g
le
with
f
in
e
-
g
r
ain
ed
s
em
an
tic
d
is
tin
ctio
n
s
an
d
lac
k
in
ter
p
r
e
tab
ilit
y
.
Sy
m
b
o
lic
m
eth
o
d
s
o
f
f
er
tr
a
n
s
p
ar
en
c
y
b
u
t
f
ail
t
o
ca
p
tu
r
e
c
o
n
tex
tu
al
n
u
an
ce
s
.
Mo
s
t
n
o
tab
ly
,
c
u
r
r
en
t
h
y
b
r
i
d
ap
p
r
o
ac
h
es
f
o
r
Fre
n
ch
eith
er
r
ely
o
n
s
tatic
co
m
b
in
atio
n
s
o
f
co
m
p
o
n
en
ts
o
r
f
ail
to
p
r
o
v
id
e
ad
ap
tiv
e
d
i
s
am
b
ig
u
atio
n
m
ec
h
an
is
m
s
,
p
ar
ticu
lar
ly
f
o
r
lex
ically
id
en
tical
b
u
t
s
em
an
tically
d
iv
er
g
en
t
s
en
te
n
ce
s
.
T
h
is
r
esear
c
h
g
a
p
is
e
s
p
ec
ially
p
r
o
n
o
u
n
ce
d
in
Fre
n
ch
,
w
h
er
e
e
x
is
tin
g
m
eth
o
d
s
ca
n
n
o
t a
d
eq
u
ately
h
a
n
d
le
th
e
lan
g
u
ag
e
’
s
c
o
m
p
lex
a
g
r
ee
m
en
t r
u
les an
d
co
n
tex
tu
al
d
ep
en
d
e
n
cies.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mea
s
u
r
in
g
s
ema
n
tic
s
imila
r
ity
in
lexi
ca
lly
id
en
tica
l
…
(
B
tis
s
a
m
E
l J
a
n
a
ti
)
955
T
h
is
s
tu
d
y
ad
d
r
ess
es th
ese
lim
itatio
n
s
th
r
o
u
g
h
th
r
ee
p
r
im
ar
y
o
b
jectiv
es:
a.
Dev
elo
p
a
d
y
n
am
ic
h
y
b
r
id
f
r
a
m
ewo
r
k
th
at
ad
ap
tiv
ely
in
teg
r
ates
s
y
m
b
o
lic
an
d
n
eu
r
al
m
et
h
o
d
s
f
o
r
Fre
n
c
h
s
em
an
tic
s
im
ilar
ity
.
b.
C
r
ea
te
an
in
ter
p
r
etab
le
d
is
am
b
ig
u
atio
n
s
y
s
tem
th
at
m
ain
tain
s
co
n
tex
tu
al
s
en
s
itiv
ity
wh
ile
p
r
o
v
id
in
g
tr
an
s
p
ar
en
t d
ec
is
io
n
-
m
ak
in
g
.
c.
E
s
tab
lis
h
a
r
o
b
u
s
t
ev
alu
atio
n
m
eth
o
d
o
lo
g
y
th
at
ac
cu
r
ately
m
ea
s
u
r
es
alig
n
m
e
n
t
with
h
u
m
an
s
em
an
tic
ju
d
g
m
en
ts
.
T
h
is
wo
r
k
m
ak
es th
r
ee
s
ig
n
if
i
ca
n
t c
o
n
tr
ib
u
tio
n
s
to
Fre
n
c
h
N
L
P:
−
An
ad
ap
tiv
e
h
y
b
r
i
d
ar
ch
itectu
r
e
th
at
s
u
r
p
ass
es th
e
lim
itatio
n
s
o
f
s
tatic
co
m
b
in
atio
n
ap
p
r
o
ac
h
es
.
−
Dem
o
n
s
tr
ated
ef
f
ec
tiv
e
n
ess
o
n
co
m
p
le
x
Fre
n
ch
a
m
b
ig
u
ities
wh
er
e
cu
r
r
en
t m
eth
o
d
s
f
ail
.
−
A
v
alid
atio
n
f
r
a
m
ewo
r
k
s
h
o
win
g
1
1
.
7
%
m
ea
n
a
b
s
o
lu
te
er
r
o
r
(
MA
E
)
r
ed
u
ctio
n
an
d
6
5
% b
et
ter
h
u
m
an
alig
n
m
en
t c
o
m
p
ar
ed
t
o
s
tate
-
of
-
th
e
-
ar
t sy
s
tem
s
.
T
h
e
r
em
ain
d
er
o
f
th
is
p
ap
er
is
o
r
g
an
ized
as
f
o
ll
o
ws:
s
ec
tio
n
2
r
e
v
iews
r
elate
d
wo
r
k
,
s
ec
tio
n
3
d
etails
o
u
r
m
eth
o
d
o
lo
g
y
,
s
ec
tio
n
4
p
r
esen
ts
ex
p
er
im
en
tal
v
alid
at
io
n
,
s
ec
tio
n
5
d
is
cu
s
s
es
f
in
d
in
g
s
,
an
d
s
ec
tio
n
6
co
n
clu
d
es with
f
u
tu
r
e
d
i
r
ec
tio
n
s
.
2.
RE
L
AT
E
D
WO
RK
R
esear
ch
o
n
s
em
an
tic
s
im
ilar
i
ty
an
d
wo
r
d
s
en
s
e
d
is
am
b
ig
u
a
tio
n
(
W
SD)
h
as
ev
o
lv
ed
f
r
o
m
s
y
m
b
o
lic
an
d
r
u
le
-
b
ased
s
y
s
tem
s
to
v
ec
to
r
-
b
ased
m
o
d
els
an
d
f
i
n
ally
to
d
ee
p
n
eu
r
al
a
n
d
m
u
ltimo
d
al
ar
ch
itectu
r
es
.
E
ac
h
p
ar
ad
ig
m
a
d
d
r
ess
es
s
em
an
tic
r
ep
r
esen
tatio
n
f
r
o
m
d
if
f
er
en
t
th
eo
r
etica
l
an
d
co
m
p
u
ta
tio
n
al
p
er
s
p
ec
tiv
es,
p
ar
ticu
lar
ly
f
o
r
Fre
n
c
h
,
wh
e
r
e
m
o
r
p
h
o
lo
g
ical
r
ich
n
ess
a
n
d
p
o
ly
s
em
y
m
ak
e
d
is
am
b
i
g
u
atio
n
esp
ec
ially
ch
allen
g
in
g
.
2
.
1
.
Cla
s
s
ica
l a
pp
ro
a
ches
E
ar
ly
m
eth
o
d
s
f
o
r
s
em
an
tic
s
im
ilar
ity
r
elied
o
n
s
y
m
b
o
lic
an
d
r
u
le
-
b
ased
f
r
am
ewo
r
k
s
s
u
ch
a
s
W
o
r
d
Net,
wh
ich
m
o
d
eled
s
y
n
tactic
an
d
s
em
an
tic
r
elatio
n
s
th
r
o
u
g
h
lo
g
ical
in
f
er
e
n
ce
.
T
h
ese
s
y
s
tem
s
wer
e
h
ig
h
ly
i
n
ter
p
r
eta
b
le
b
u
t
f
ac
e
d
m
ajo
r
ch
allen
g
es:
m
an
u
al
r
u
le
d
ep
en
d
en
ce
,
p
o
o
r
s
ca
l
ab
ilit
y
,
an
d
lim
ited
lin
g
u
is
tic
ad
ap
tab
ilit
y
[
1
]
.
I
n
Ar
ab
ic
NL
P,
f
o
r
e
x
am
p
l
e,
E
n
aa
n
ai
et
a
l.
[
2
]
p
r
o
p
o
s
ed
a
m
o
r
p
h
o
s
em
an
tic
f
ilter
i
n
g
m
eth
o
d
ad
d
r
ess
in
g
am
b
ig
u
ity
ca
u
s
ed
b
y
a
g
g
lu
tin
atio
n
an
d
lac
k
o
f
v
o
ca
lizatio
n
.
T
h
eir
s
y
s
tem
ex
p
a
n
d
ed
wo
r
d
s
in
to
all
p
o
s
s
ib
le
v
o
ca
lized
f
o
r
m
s
a
n
d
d
is
am
b
ig
u
ated
m
ea
n
in
g
s
u
s
in
g
co
n
te
x
tu
al
r
u
les,
a
u
s
er
p
r
o
f
i
le,
an
d
th
e
s
em
an
tic
lex
ico
n
AR
R
AM
OUZ
AL
W
A
SEE
T
,
with
a
J
ac
ca
r
d
-
b
ased
s
im
ilar
ity
f
u
n
ctio
n
t
o
en
h
an
ce
alig
n
m
en
t.
T
h
o
u
g
h
ef
f
ec
tiv
e,
th
is
ap
p
r
o
ac
h
r
em
ai
n
ed
co
n
s
tr
ain
ed
b
y
r
u
le
c
o
m
p
l
ex
ity
an
d
lim
ited
p
o
r
ta
b
ilit
y
to
o
th
er
lan
g
u
ag
es.
Su
b
s
eq
u
en
t
v
ec
t
o
r
-
b
ased
m
o
d
els,
in
clu
d
in
g
b
a
g
-
of
-
wo
r
d
s
(
B
o
W
)
an
d
T
F
-
I
DF
[
1
]
,
p
r
o
v
id
ed
q
u
an
titativ
e
tex
t
r
ep
r
esen
tati
o
n
s
b
u
t
ig
n
o
r
ed
s
em
an
tic
r
e
latio
n
s
h
ip
s
.
W
o
r
d
em
b
ed
d
i
n
g
m
o
d
els
im
p
r
o
v
ed
lex
ical
g
en
er
aliza
tio
n
Fas
tTe
x
t
[
3
]
,
[
4
]
in
tr
o
d
u
ce
d
s
u
b
w
o
r
d
in
f
o
r
m
atio
n
,
wh
ile
W
o
r
d
2
Vec
[
5
]
an
d
Glo
Ve
[
6
]
ca
p
tu
r
ed
co
-
o
cc
u
r
r
en
ce
p
atter
n
s
.
Ho
wev
er
,
th
ese
m
o
d
els
wer
e
co
n
tex
t
-
in
s
en
s
itiv
e
an
d
s
tr
u
g
g
led
with
p
o
ly
s
em
y
.
L
ater
p
r
o
b
a
b
ilis
tic
em
b
ed
d
in
g
s
,
s
u
ch
as
Ar
o
r
a
et
a
l.
[
7
]
,
e
n
h
an
ce
d
g
l
o
b
al
s
tr
u
c
tu
r
e
b
u
t
s
till
r
elied
o
n
s
im
p
lify
in
g
Gau
s
s
ian
ass
u
m
p
tio
n
s
.
T
h
ese
lim
itatio
n
s
m
o
tiv
ated
t
h
e
em
er
g
e
n
ce
o
f
co
n
tex
tu
ali
ze
d
tr
an
s
f
o
r
m
er
-
b
ased
m
o
d
el
s
s
u
ch
as
B
E
R
T
[
8
]
,
wh
ich
in
tr
o
d
u
ce
d
d
y
n
am
ic
em
b
ed
d
in
g
s
s
en
s
itiv
e
to
s
u
r
r
o
u
n
d
in
g
wo
r
d
s
,
i
m
p
r
o
v
i
n
g
s
em
an
tic
co
h
er
en
ce
b
u
t r
aisi
n
g
is
s
u
es o
f
co
m
p
u
tatio
n
al
co
s
t a
n
d
in
te
r
p
r
etab
ilit
y
.
2
.
2
.
P
re
t
ra
ined la
ng
ua
g
e
m
o
dels
Pre
tr
ain
ed
lan
g
u
ag
e
m
o
d
els
(
P
L
Ms)
b
ased
o
n
t
h
e
t
r
an
s
f
o
r
m
e
r
ar
ch
itectu
r
e
h
av
e
t
r
an
s
f
o
r
m
e
d
NL
P
b
y
en
ab
lin
g
c
o
n
tex
t
-
awa
r
e
a
n
d
m
u
ltil
in
g
u
al
r
ep
r
esen
tatio
n
s
.
Fo
r
Fre
n
ch
,
n
o
tab
le
m
o
d
els
in
clu
d
e
C
am
em
B
E
R
T
[
6
]
,
a
R
o
B
E
R
T
a
-
b
ased
m
o
d
el
tr
ain
ed
o
n
OSC
AR
an
d
Fre
n
ch
W
ik
ip
ed
ia;
Flau
B
E
R
T
[
1
]
,
a
lar
g
e
1
2
–
24
-
lay
er
m
o
d
el
with
r
ich
le
x
ical
co
v
er
ag
e;
B
E
R
T
wee
tFR
[
9
]
-
[
1
2
]
,
s
p
ec
ialized
f
o
r
T
witter
;
an
d
m
B
E
R
T
[
8
]
a
m
u
ltil
in
g
u
al
m
o
d
el
less
o
p
tim
ized
f
o
r
Fre
n
ch
s
y
n
tax
an
d
m
o
r
p
h
o
lo
g
y
.
Sem
an
tic
s
im
ilar
ity
is
u
s
u
ally
co
m
p
u
ted
f
r
o
m
th
e
[
C
L
S]
to
k
en
o
r
m
ea
n
-
p
o
o
led
em
b
ed
d
in
g
s
u
s
in
g
co
s
in
e
s
im
ilar
ity
.
T
o
b
etter
ca
p
tu
r
e
s
en
ten
ce
-
lev
el
m
ea
n
in
g
,
Sen
t
en
ce
-
B
E
R
T
(
SB
E
R
T
)
in
tr
o
d
u
ce
d
a
Siam
ese
ar
ch
itectu
r
e,
later
ad
ap
ted
f
o
r
Fre
n
ch
in
C
am
em
B
E
R
T
-
Sen
ten
ce
an
d
L
aBS
E
[
13
]
.
B
en
ch
m
ar
k
s
s
u
ch
as
STS
-
FR
p
r
o
v
id
e
e
v
alu
atio
n
d
atasets
,
th
o
u
g
h
th
ey
ar
e
lim
it
ed
i
n
d
o
m
ain
d
iv
er
s
ity
.
Desp
ite
th
eir
s
u
cc
ess
,
PLM
s
s
t
ill
f
a
ce
ch
allen
g
es
with
m
o
r
p
h
o
lo
g
ical
v
a
r
iatio
n
,
r
e
g
i
s
ter
d
iv
er
s
ity
,
an
d
b
iases
f
r
o
m
lim
ited
Fre
n
ch
co
r
p
o
r
a
,
wh
ich
ca
n
lead
to
s
em
an
tic
d
r
if
t
in
p
o
ly
s
em
o
u
s
co
n
tex
ts
,
h
ig
h
lig
h
tin
g
th
e
n
ee
d
f
o
r
m
o
r
e
r
o
b
u
s
t
a
n
d
co
n
tex
t
-
s
en
s
itiv
e
ap
p
r
o
ac
h
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
41
,
No
.
3
,
Ma
r
ch
20
2
6
:
9
5
4
-
965
956
2
.
3
.
Co
ntr
a
s
t
iv
e
lea
rning
a
n
d su
perv
i
s
ed
m
et
ho
ds
R
ec
en
t
ap
p
r
o
ac
h
es
h
av
e
ad
v
an
ce
d
b
ey
o
n
d
s
tatic
em
b
ed
d
in
g
s
b
y
in
co
r
p
o
r
atin
g
c
o
n
tr
a
s
tiv
e
an
d
g
en
er
ativ
e
lear
n
i
n
g
f
r
a
m
ewo
r
k
s
.
SimCS
E
[
14
]
r
ef
in
es
s
en
ten
ce
r
ep
r
esen
tatio
n
s
th
r
o
u
g
h
c
o
n
tr
asti
v
e
lear
n
in
g
,
u
s
in
g
d
r
o
p
o
u
t
-
in
d
u
ce
d
p
o
s
itiv
es
in
u
n
s
u
p
e
r
v
is
ed
m
o
d
e
an
d
p
ar
a
p
h
r
ase
p
air
s
with
n
eg
a
tiv
es
in
s
u
p
er
v
is
ed
m
o
d
e.
Fre
n
ch
ad
a
p
tatio
n
s
f
in
e
-
tu
n
ed
o
n
FrSem
L
ex
im
p
r
o
v
e
d
STS
-
FR
p
er
f
o
r
m
an
ce
an
d
s
em
an
tic
p
r
ec
is
io
n
in
s
en
ten
ce
alig
n
m
en
t.
T
5
[
1
5
]
an
d
its
Fre
n
ch
v
ar
ia
n
t
ST5
[
16
]
,
[
1
7
]
r
ec
o
n
ce
p
t
u
al
ize
s
im
ilar
ity
as
a
tex
t
g
en
er
atio
n
task
,
p
r
o
d
u
cin
g
r
ef
o
r
m
u
latio
n
s
o
r
s
im
ilar
ity
s
co
r
es
th
r
o
u
g
h
m
u
ltit
ask
lear
n
in
g
.
E
m
p
ir
ical
r
esu
lts
s
h
o
w
SimC
SE
[
2
]
-
R
o
B
E
R
T
a
-
lar
g
e
(
2
0
2
1
)
r
ea
c
h
in
g
8
6
.
7
o
n
STS
-
B
,
wh
ile
S
T
5
an
d
m
ST5
ac
h
iev
e
u
p
to
ρ
=
8
3
.
2
o
n
XSTS,
co
n
f
ir
m
in
g
th
e
c
o
m
p
lem
en
ta
r
ity
b
etwe
en
co
n
tr
asti
v
e
an
d
g
e
n
er
ativ
e
p
ar
a
d
ig
m
s
f
o
r
s
em
an
t
ic
s
im
ilar
ity
task
s
.
2
.
4
.
H
y
brid a
nd
m
ultim
o
da
l
a
pp
ro
a
ches f
o
r
a
dv
a
nced
N
L
P
Hy
b
r
id
an
d
m
u
ltimo
d
al
ar
c
h
itectu
r
es
co
m
b
in
e
s
y
m
b
o
lic
in
ter
p
r
etab
ilit
y
with
d
ee
p
co
n
tex
tu
al
u
n
d
er
s
tan
d
i
n
g
.
Sy
s
tem
s
s
u
ch
as
Sem
Glo
Ve
[
18
]
,
[
1
9
]
in
teg
r
ate
Glo
Ve
’
s
g
lo
b
al
co
-
o
cc
u
r
r
en
ce
s
tatis
tic
s
with
B
E
R
T
’
s
co
n
tex
tu
al
em
b
ed
d
in
g
s
,
ac
h
iev
in
g
1
2
% g
ain
s
in
le
x
ical
s
im
ilar
ity
an
d
im
p
r
o
v
ed
W
SD
.
C
r
o
s
s
-
m
o
d
al
m
o
d
els
lik
e
Sim
-
C
L
I
P
[
2
0
]
,
DiC
A
[
2
1
]
,
an
d
MCS
E
[
2
2
]
ex
ten
d
th
ese
p
r
in
cip
les
to
v
is
io
n
–
lan
g
u
a
g
e
a
n
d
a
u
d
io
–
tex
t
d
o
m
ain
s
,
u
s
in
g
co
n
t
r
asti
v
e
alig
n
m
en
t
f
o
r
p
a
r
tially
lab
el
ed
o
r
u
n
an
n
o
tated
d
atasets
.
R
ec
en
t
ad
v
an
ce
s
in
m
u
ltil
in
g
u
al
m
u
ltimo
d
al
m
o
d
elin
g
,
s
u
ch
as
m
u
ltil
in
g
u
al
Sen
ten
ce
-
T
5
[
2
3
]
,
en
s
u
r
e
cr
o
s
s
-
lin
g
u
al
s
em
a
n
tic
co
n
s
is
ten
cy
an
d
en
ab
le
c
r
o
s
s
-
m
o
d
al
d
is
am
b
ig
u
atio
n
,
b
en
ef
itin
g
m
u
ltil
in
g
u
al
tr
an
s
latio
n
an
d
m
u
ltimo
d
al
r
et
r
iev
al.
Desp
ite
th
ese
ac
h
iev
em
en
ts
,
m
o
s
t
s
tu
d
ies
f
o
cu
s
o
n
E
n
g
lis
h
,
with
lim
ited
ex
p
lo
r
atio
n
in
Fre
n
ch
esp
ec
ially
f
o
r
lex
ically
id
en
tical
b
u
t
s
em
an
tically
d
iv
er
g
e
n
t
s
en
ten
ce
s
,
wh
ich
r
em
ain
a
cr
itical
ch
allen
g
e
f
o
r
co
n
te
x
t
-
s
en
s
itiv
e
u
n
d
er
s
tan
d
in
g
.
2
.
5
.
Co
m
pa
ra
t
iv
e
s
um
ma
ry
o
f
ma
j
o
r
m
o
del f
a
m
ilies
T
ab
le
1
s
u
m
m
ar
izes
m
ajo
r
NL
P
m
o
d
el
f
am
ilies
,
h
ig
h
lig
h
tin
g
r
e
p
r
esen
tativ
e
m
o
d
els,
k
ey
ad
v
an
ce
m
e
n
ts
,
lim
itatio
n
s
,
an
d
r
ep
o
r
ted
p
er
f
o
r
m
an
ce
to
p
r
o
v
id
e
a
co
n
cise
co
m
p
ar
ativ
e
o
v
er
v
iew
o
f
cu
r
r
en
t
ap
p
r
o
ac
h
es.
T
ab
le
1
.
NL
P m
o
d
els f
am
ilies
:
f
ea
tu
r
es,
lim
itatio
n
s
,
an
d
p
er
f
o
r
m
an
ce
s
M
o
d
e
l
f
a
mi
l
y
R
e
p
r
e
se
n
t
a
t
i
v
e
m
o
d
e
l
s
K
e
y
f
e
a
t
u
r
e
s
/
p
r
o
g
r
e
ss
Li
mi
t
a
t
i
o
n
s
P
e
r
f
o
r
ma
n
c
e
/
N
o
t
e
s
M
o
d
e
l
f
a
mi
l
y
V
e
c
t
o
r
-
b
a
s
e
d
LSA
,
W
o
r
d
2
V
e
c
,
G
l
o
V
e
,
A
r
o
r
a
Ef
f
i
c
i
e
n
t
e
mb
e
d
d
i
n
g
s,
l
a
t
e
n
t
/
g
l
o
b
a
l
s
e
ma
n
t
i
c
s
N
o
c
o
n
t
e
x
t
,
i
g
n
o
r
e
s
w
o
r
d
o
r
d
e
r
,
m
e
m
o
r
y
-
h
e
a
v
y
–
V
e
c
t
o
r
-
b
a
s
e
d
Tr
a
n
sf
o
r
mer
P
LM
s
B
ER
T,
F
l
a
u
B
ER
T
,
S
B
ER
T,
C
a
me
mBE
R
T
,
B
ER
Tw
e
e
t
F
R
C
o
n
t
e
x
t
u
a
l
e
mb
e
d
d
i
n
g
s,
se
n
t
e
n
c
e
-
l
e
v
e
l
p
o
o
l
i
n
g
H
i
g
h
c
o
s
t
,
l
o
w
e
x
p
l
a
i
n
a
b
i
l
i
t
y
M
N
LI
:
8
4
.
6
,
S
TS
-
B
:
8
4
.
9
,
Te
x
t
C
l
f
:
9
5
%,
F
1
:
7
1
.
2
7
Tr
a
n
sf
o
r
mer
P
LM
s
C
o
n
t
r
a
st
i
v
e
/
G
e
n
e
r
a
t
i
v
e
S
i
mCSE
,
T
5
,
m
-
S
T5
C
o
n
t
r
a
st
i
v
e
l
e
a
r
n
i
n
g
,
g
e
n
e
r
a
t
i
v
e
m
o
d
e
l
i
n
g
N
e
e
d
s
f
i
n
e
-
t
u
n
i
n
g
,
d
a
t
a
-
d
e
p
e
n
d
e
n
t
S
TS
-
B
:
7
6
.
8
5
–
8
6
.
7
0
,
G
LU
E:
9
0
.
3
,
X
S
TS ρ
:
8
3
.
2
C
o
n
t
r
a
st
i
v
e
/
G
e
n
e
r
a
t
i
v
e
H
y
b
r
i
d
/
M
u
l
t
i
m
o
d
a
l
S
e
mG
l
o
V
e
,
M
C
S
E,
D
i
C
A
,
S
i
m
-
C
LI
P
4
Emb
e
d
d
i
n
g
f
u
s
i
o
n
,
mu
l
t
i
m
o
d
a
l
l
e
a
r
n
i
n
g
C
o
m
p
u
t
a
t
i
o
n
a
l
l
y
h
e
a
v
y
A
c
c
u
r
a
c
y
:
9
5
.
6
8
%,
mA
P
:
0
.
6
9
7
,
C
I
D
Er
:
8
1
.
6
H
y
b
r
i
d
/
M
u
l
t
i
m
o
d
a
l
2
.
6
.
Dis
cus
s
io
n a
nd
ide
ntif
ie
d r
esea
rc
h g
a
p
Alth
o
u
g
h
Fre
n
c
h
PLM
s
s
u
ch
as
C
am
em
B
E
R
T
an
d
Flau
B
E
R
T
h
av
e
ac
h
iev
ed
n
o
tab
le
p
r
o
g
r
ess
,
th
ey
s
till
s
tr
u
g
g
le
with
p
o
ly
s
em
y
an
d
s
em
an
tic
am
b
i
g
u
ity
at
t
h
e
s
en
ten
ce
lev
el.
E
x
is
tin
g
b
e
n
ch
m
ar
k
s
,
s
u
c
h
as
STS
-
FR
an
d
FrSem
L
ex
,
r
em
ain
lim
ited
an
d
lack
ex
am
p
les
wh
er
e
lex
ical
id
en
tity
h
id
es
d
iv
er
g
en
t
m
ea
n
i
n
g
s
(
e.
g
.
,
f
ils
,
co
u
r
an
t,
b
r
an
c
h
e)
.
Mo
r
eo
v
er
,
h
y
b
r
id
f
r
a
m
ewo
r
k
s
co
m
b
in
in
g
s
y
m
b
o
lic
in
ter
p
r
etab
ilit
y
an
d
n
eu
r
al
r
ea
s
o
n
in
g
r
e
m
ain
u
n
d
er
e
x
p
lo
r
e
d
f
o
r
Fre
n
c
h
.
R
e
ce
n
t
m
u
ltil
in
g
u
al
m
o
d
els,
s
u
c
h
as
m
DeBERTa
[
9
]
,
L
L
aM
A
[
18
]
,
an
d
XSI
M2
3
[
1
]
,
s
h
o
w
p
r
o
m
is
in
g
p
o
ten
tia
l
f
o
r
m
o
d
elin
g
c
r
o
s
s
-
lin
g
u
al
s
em
an
tic
am
b
ig
u
ity
,
y
et
th
eir
ef
f
ec
tiv
en
ess
f
o
r
co
n
tex
tu
al
d
is
am
b
ig
u
atio
n
i
n
Fre
n
ch
r
em
ain
s
to
b
e
in
v
esti
g
a
ted
.
Ou
r
p
r
o
p
o
s
ed
h
y
b
r
id
m
eth
o
d
d
ir
ec
tly
a
d
d
r
ess
es
lex
icall
y
id
en
tical
y
et
s
em
an
tically
am
b
ig
u
o
u
s
Fre
n
ch
s
en
ten
ce
s
,
b
r
id
g
in
g
th
e
g
ap
b
etwe
en
h
u
m
an
-
lik
e
in
ter
p
r
etab
ilit
y
an
d
th
e
co
n
te
x
tu
al
ac
cu
r
ac
y
o
f
d
ee
p
lear
n
in
g
r
e
p
r
esen
tatio
n
s
.
T
h
e
n
o
v
elty
o
f
o
u
r
ap
p
r
o
ac
h
lie
s
in
its
co
n
tex
t
-
awa
r
e
f
u
s
io
n
m
ec
h
an
is
m
,
wh
ich
d
y
n
am
ically
ad
ju
s
ts
th
e
in
ter
p
lay
b
etwe
en
s
y
m
b
o
lic
an
d
n
e
u
r
al
elem
en
ts
ac
co
r
d
in
g
t
o
lin
g
u
is
tic
co
m
p
lex
ity
.
I
n
co
n
tr
ast
to
tr
ad
itio
n
al
h
y
b
r
id
m
o
d
els
r
ely
in
g
o
n
f
ix
ed
w
eig
h
tin
g
s
,
o
u
r
f
r
am
ewo
r
k
co
n
tin
u
o
u
s
ly
ass
ess
e
s
am
b
ig
u
ity
cu
es
to
f
in
e
-
t
u
n
e
th
e
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ec
is
io
n
an
d
tr
an
s
p
ar
en
c
y
.
T
h
is
s
ig
n
if
i
es
a
tr
an
s
f
o
r
m
ativ
e
s
h
if
t
f
r
o
m
r
ig
id
ar
ch
itectu
r
es
to
an
a
d
ap
tiv
e
s
y
s
tem
tailo
r
e
d
to
th
e
s
u
b
tleties
o
f
Fre
n
ch
,
ef
f
ec
tiv
ely
tack
lin
g
p
er
s
is
ten
t iss
u
es lik
e
h
o
m
o
n
y
m
y
an
d
s
tr
u
ctu
r
al
am
b
i
g
u
ity
t
h
at
ch
al
len
g
e
c
u
r
r
en
t
PLM
s
an
d
h
y
b
r
id
s
o
lu
tio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mea
s
u
r
in
g
s
ema
n
tic
s
imila
r
ity
in
lexi
ca
lly
id
en
tica
l
…
(
B
tis
s
a
m
E
l J
a
n
a
ti
)
957
3.
M
E
T
H
O
D
3.
1
.
I
ns
t
rum
ent
s
Ou
r
h
y
b
r
id
ap
p
r
o
ac
h
u
s
es
th
e
s
en
ten
ce
-
tr
an
s
f
o
r
m
er
s
/all
-
m
p
n
et
-
b
ase
-
v2
v
er
s
io
n
o
f
SB
E
R
T
,
g
en
er
atin
g
7
6
8
-
d
im
en
s
io
n
al
em
b
ed
d
in
g
s
.
E
x
p
er
im
en
ts
w
er
e
co
n
d
u
cted
o
n
Go
o
g
le
C
o
lab
in
a
C
PU
en
v
ir
o
n
m
en
t
with
th
e
f
o
llo
wi
n
g
p
ar
am
eter
s
:
b
atch
s
ize
o
f
3
2
,
tem
p
er
at
u
r
e
o
f
0
.
0
5
f
o
r
th
e
So
f
tMa
x
f
u
n
ctio
n
,
an
d
s
im
ilar
ity
th
r
esh
o
ld
o
f
0
.
6
.
T
h
e
im
p
lem
e
n
tatio
n
in
teg
r
ates
W
o
r
d
Net
f
o
r
s
y
n
s
et
ex
tr
ac
tio
n
,
NL
T
K
an
d
Sp
aCy
f
o
r
lin
g
u
is
tic
p
r
ep
r
o
ce
s
s
in
g
,
an
d
cu
s
to
m
r
o
u
tin
es
f
o
r
ca
lcu
latin
g
weig
h
ted
f
u
zz
y
J
ac
ca
r
d
.
T
h
e
f
u
s
io
n
m
ec
h
an
is
m
d
y
n
am
ically
co
m
b
in
es
co
s
in
e
s
im
ilar
itie
s
(
wit
h
L
2
n
o
r
m
aliza
tio
n
)
an
d
s
y
m
b
o
li
c
s
im
ilar
itie
s
th
r
o
u
g
h
an
a
d
ap
tiv
e
atten
tio
n
f
u
n
ctio
n
b
ased
o
n
d
etec
ted
lin
g
u
is
tic
co
m
p
lex
ity
.
3
.
2
.
O
rg
a
ni
g
ra
m
m
e
T
h
e
h
y
b
r
i
d
m
eth
o
d
co
m
b
in
es
lin
g
u
is
tic
d
is
am
b
ig
u
atio
n
an
d
s
em
an
tic
v
ec
to
r
m
o
d
elin
g
th
r
o
u
g
h
eig
h
t
s
eq
u
en
tial step
s
s
u
m
m
ar
ized
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
W
o
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
a
p
p
r
o
ac
h
3
.
3
.
Ste
p
1:
t
o
k
en
f
ilt
e
ring
a
nd
a
m
big
uity
det
ec
t
io
n
E
ac
h
to
k
en
in
th
e
s
en
ten
ce
is
ch
ec
k
ed
ag
ain
s
t
th
e
W
o
r
d
Net
lex
ical
d
atab
ases
to
d
etec
t
am
b
ig
u
o
u
s
ter
m
s
,
wh
ich
f
o
r
m
th
e
s
et
A,
wh
ile
u
n
am
b
ig
u
o
u
s
wo
r
d
s
f
o
r
m
th
e
s
et
C
.
Fo
r
e
x
am
p
le,
in
t
h
e
s
en
ten
ce
“
L
es
f
ils
d
e
co
n
d
u
cte
u
r
o
n
t
co
u
p
é
le
c
o
u
r
an
t
”
(
“
T
h
e
d
r
iv
er
’
s
s
o
n
s
c
u
t
th
e
p
o
wer
”
)
,
we
o
b
tain
:
A
=
{f
ils
,
co
n
d
u
cteu
r
}
an
d
C
=
{o
n
t,
c
o
u
p
é
,
co
u
r
an
t}
.
Sto
p
wo
r
d
s
ar
e
r
em
o
v
e
d
,
an
d
lem
m
atiza
tio
n
is
ap
p
lied
to
s
tan
d
ar
d
ize
th
e
lex
ic
al
f
o
r
m
s
.
T
h
is
p
r
ep
r
o
ce
s
s
in
g
s
tep
en
s
u
r
es th
at
th
e
s
im
ilar
ity
ca
lcu
latio
n
s
.
3
.
4
.
Ste
p
2:
l
ex
ica
l sens
e
(
Wo
rdnet
)
3
.
4
.
1
.
P
o
ly
s
emo
us
t
er
m
s
Fo
r
ea
ch
p
o
l
y
s
em
o
u
s
ter
m
,
E
n
g
lis
h
d
ef
in
itio
n
s
wer
e
r
etr
i
ev
ed
f
r
o
m
W
o
r
d
Net
an
d
tr
an
s
lated
in
to
Fre
n
ch
,
r
etain
in
g
th
eir
u
n
iq
u
e
s
y
n
s
et
id
en
tifie
r
s
.
“
f
ils
”
:
v
11
:
“
u
n
co
n
d
u
cteu
r
m
étalliq
u
e
q
u
i
tr
an
s
p
o
r
te
l
’
élec
tr
icité
s
u
r
u
n
e
d
is
tan
ce
”
(
wir
e.
n
.
0
2
)
.
v
12
:
“
u
n
e
p
r
o
g
é
n
itu
r
e
h
u
m
ain
e
m
a
s
cu
lin
e
”
(
s
o
n
.
n
.
0
1
)
.
“
co
n
d
u
cte
u
r
”
:
v
21
:
“
l
’
o
p
ér
ateu
r
d
’
u
n
v
éh
ic
u
le
à
m
o
te
u
r
”
(
d
r
iv
er
.
n
.
0
1
)
.
v
22
:
“
u
n
ap
p
ar
eil
co
n
ç
u
p
o
u
r
tr
a
n
s
m
ettr
e
l
’
élec
tr
icité,
la
ch
aleu
r
,
etc.
”
(
co
n
d
u
cto
r
.
n
.
0
4
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
41
,
No
.
3
,
Ma
r
ch
20
2
6
:
9
5
4
-
965
958
3
.
4
.
2
.
G
lo
s
s
prepro
ce
s
s
ing
Glo
s
s
es
ar
e
p
r
ep
r
o
ce
s
s
ed
b
y
r
em
o
v
in
g
Fre
n
ch
s
to
p
wo
r
d
s
,
lem
m
atizin
g
n
o
u
n
s
an
d
v
er
b
s
,
an
d
d
eletin
g
o
cc
u
r
r
en
ce
s
o
f
th
e
tar
g
et
wo
r
d
to
a
v
o
id
b
ias.
E
x
a
m
p
le
:
co
n
d
u
cte
u
r
(
co
n
d
u
ct
o
r
.
n
.
0
4
)
o
r
ig
in
al
g
lo
s
s
:
“
u
n
ap
p
ar
eil
co
n
ç
u
p
o
u
r
tr
a
n
s
m
ettr
e
l
’
élec
tr
icité,
la
ch
al
eu
r
,
etc.
”
Af
ter
p
r
ep
r
o
ce
s
s
in
g
:
“
ap
p
ar
eil
co
n
ce
v
o
i
r
tr
an
s
m
ettr
e
élec
tr
icité
ch
aleu
r
”
.
T
h
is
y
ield
s
clea
n
,
s
tan
d
ar
d
ized
g
lo
s
s
es
f
o
r
p
r
ec
is
e
s
em
an
tic
s
im
ilar
ity
co
m
p
u
tatio
n
.
3
.
5
.
Ste
p
3:
h
y
brid co
nte
x
t
u
a
l sim
ila
rit
y
m
esu
re
m
ent
Sem
an
tic
co
h
er
en
ce
b
etwe
en
ca
n
d
id
ate
s
en
s
es
o
f
am
b
ig
u
o
u
s
wo
r
d
s
an
d
th
eir
co
n
te
x
tu
al
ter
m
s
is
ev
alu
ated
u
s
in
g
a
h
y
b
r
id
m
o
d
e
l in
teg
r
atin
g
weig
h
te
d
f
u
zz
y
ja
cc
ar
d
an
d
SB
E
R
T
em
b
ed
d
in
g
s
.
3
.
5
.
1
.
Weig
hte
d
f
uzzy
j
a
cc
a
rd
T
h
e
weig
h
ted
f
u
zz
y
jacc
a
r
d
,
ad
ap
ted
f
o
r
Fre
n
ch
NL
P,
m
ea
s
u
r
es
lex
ical
o
v
e
r
lap
b
etwe
en
g
lo
s
s
es
o
f
ca
n
d
id
ate
s
en
s
es
an
d
c
o
n
tex
t
wo
r
d
s
’
d
o
m
in
an
t
s
en
s
es.
Sco
r
es
r
an
g
e
f
r
o
m
0
to
1
,
with
1
in
d
icatin
g
p
er
f
ec
t
alig
n
m
en
t.
J
fw
(
A
,
B
)
=
∑
∑
j
∈
B
i
∈
A
(
w
i
,
w
j
)
.
s
(
i
,
j
)
∑
w
i
i
∈
A
+
∑
w
j
j
∈
B
−
∑
∑
j
∈
B
i
∈
A
(
w
i
,
w
j
)
.
s
(
i
,
j
)
(
1
)
w
i
,
w
j
:
ter
m
weig
h
ts
.
A
,
B
:
p
r
ep
r
o
ce
s
s
ed
g
lo
s
s
es o
f
th
e
tar
g
et
wo
r
d
s
en
s
e
an
d
co
n
tex
t
wo
r
d
d
o
m
in
an
t sen
s
e
.
s
(
i
,
j
)
:
s
em
an
tic
s
im
ilar
ity
b
etwe
en
ter
m
s
i a
n
d
j
.
Ou
tp
u
t:
T
h
e
weig
h
ted
f
u
z
zy
J
ac
ca
r
d
i
n
d
ex
,
ad
a
p
ted
f
o
r
Fre
n
ch
NL
P,
m
ea
s
u
r
es
lex
ical
o
v
er
lap
b
etwe
en
p
r
ep
r
o
ce
s
s
ed
g
lo
s
s
es
o
f
ca
n
d
id
ate
s
en
s
es
an
d
th
e
d
o
m
i
n
an
t
co
n
tex
t
s
e
n
s
es.
E
ac
h
p
air
is
s
co
r
ed
f
r
o
m
0
to
1
,
with
1
in
d
icatin
g
p
er
f
ec
t sem
an
tic
alig
n
m
en
t.
Pre
p
r
o
ce
s
s
in
g
:
−
L
em
m
atiza
tio
n
(
“
co
u
p
és
”
⇾
“
co
u
p
er
”
)
−
Sto
p
wo
r
d
r
em
o
v
al
(
“
de
”
,
“
le
”
,
“
ont
”
)
−
B
ias m
itig
atio
n
(
ex
clu
d
e
tar
g
e
t w
o
r
d
f
r
o
m
o
w
n
g
lo
s
s
)
3
.
5
.
2
.
Co
s
ine
-
ba
s
ed
s
em
a
ntic
a
na
ly
s
is
f
o
r
F
re
nch
us
ing
m
ultiling
ua
l S
B
E
R
T
T
o
ca
p
tu
r
e
d
ee
p
er
s
em
an
tic
r
elatio
n
s
h
ip
s
b
ey
o
n
d
lex
ical
o
v
er
lap
,
m
u
ltil
in
g
u
al
SB
E
R
T
em
b
ed
d
in
g
s
co
m
p
u
te
co
s
in
e
s
im
ilar
ity
b
etwe
en
s
en
ten
ce
v
ec
to
r
s
in
h
ig
h
-
d
im
e
n
s
io
n
al
s
p
ac
e,
ef
f
ec
tiv
ely
m
o
d
elin
g
s
y
n
tactic
an
d
co
n
ce
p
tu
al
d
e
p
e
n
d
en
cies,
with
o
p
tim
izatio
n
f
o
r
Fre
n
ch
.
(
,
)
=
(
)
∈
[
0
,
1
]
(
2
)
:
I
n
p
u
t te
x
t
r
e
p
r
esen
tatio
n
.
:
an
g
le
b
etwe
en
em
b
ed
d
i
n
g
v
e
cto
r
s
in
h
ig
h
d
im
en
s
io
n
al
s
p
ac
e
.
3
.
5
.
3
.
H
y
brid s
co
ring
f
o
r
m
ula
T
h
e
co
n
tex
t
u
al
s
im
ilar
ity
b
etwe
en
ca
n
d
id
ate
w
o
r
d
s
an
d
th
eir
s
u
r
r
o
u
n
d
in
g
c
o
n
tex
t w
as e
v
alu
ated
u
s
in
g
a
h
y
b
r
id
m
o
d
el
th
at
c
o
m
b
in
es
weig
h
ted
f
u
zz
y
jacc
ar
d
an
d
SB
E
R
T
em
b
ed
d
in
g
s
(
T
a
b
le
2
)
:
=
×
+
(
1
−
)
×
(
3
)
Par
am
eter
s
:
−
α
=0
.
7
:
w
eig
h
t f
o
r
lex
ical
s
im
ilar
ity
u
s
in
g
weig
h
te
d
f
u
zz
y
jac
ca
r
d
.
−
1
-
α
=0
.
3
:
w
eig
h
t f
o
r
s
em
an
tic
s
im
ilar
ity
(
SB
E
R
T
)
.
T
h
e
weig
h
tin
g
p
ar
am
ete
r
(
α
)
co
n
tr
o
ls
th
e
r
elativ
e
co
n
tr
ib
u
ti
o
n
o
f
lex
ical
s
im
ilar
ity
(
weig
h
ted
f
u
zz
y
jacc
ar
d
)
an
d
s
em
an
tic
s
im
i
lar
ity
(
SB
E
R
T
)
.
T
o
d
eter
m
in
e
t
h
e
o
p
tim
al
v
alu
e
o
f
α
,
a
Py
th
o
n
-
b
ased
p
r
o
g
r
am
test
ed
v
alu
es
r
an
g
in
g
f
r
o
m
0
to
1
in
in
cr
e
m
en
ts
o
f
0
.
1
,
u
s
in
g
s
en
ten
ce
p
air
s
th
at
a
r
e
l
ex
ically
s
im
ilar
b
u
t
s
em
an
tically
d
iv
er
g
en
t.
T
h
e
p
r
o
g
r
am
a
u
to
m
atica
lly
s
elec
ted
th
e
α
v
alu
e
t
h
at
m
ax
im
ized
t
h
e
o
v
er
all
s
im
ilar
ity
ac
cu
r
ac
y
,
r
esu
ltin
g
in
a
n
o
p
ti
m
al
α
=
0
.
7
.
T
h
is
co
n
f
ig
u
r
atio
n
p
r
io
r
itizes
lex
ical
s
im
ilar
it
y
wh
ile
p
r
eser
v
in
g
s
em
an
tic
co
n
tex
tu
al
n
u
an
ce
s
,
ac
h
iev
in
g
a
b
alan
ce
d
p
e
r
f
o
r
m
an
ce
in
d
etec
tin
g
s
em
an
tic
d
iv
er
g
en
ce
s
am
o
n
g
lex
ical
ly
clo
s
e
s
en
ten
ce
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mea
s
u
r
in
g
s
ema
n
tic
s
imila
r
ity
in
lexi
ca
lly
id
en
tica
l
…
(
B
tis
s
a
m
E
l J
a
n
a
ti
)
959
T
ab
le
2
.
Hy
b
r
id
c
o
n
tex
tu
al
s
i
m
ilar
ity
m
e
a
s
u
r
em
en
t
Ta
r
g
e
t
w
o
r
d
S
B
ER
T
(
c
o
u
p
é
)
S
B
ER
T
(
c
o
u
r
a
n
t
)
Jac
c
a
r
d
(
c
o
u
p
é
)
Jac
c
a
r
d
(
c
o
u
r
a
n
t
)
G
l
o
b
a
l
s
c
o
r
e
F
i
l
s
0
.
2
8
8
0
.
5
7
2
0
.
5
7
0
0
.
7
0
8
0
.
5
3
4
F
i
l
s
0
.
4
3
8
0
.
2
8
8
0
.
5
9
2
0
.
4
4
1
0
.
4
3
9
C
o
n
d
u
c
t
e
u
r
0
.
4
9
6
0
.
3
7
2
0
.
8
4
2
0
.
7
3
0
0
.
6
0
9
C
o
n
d
u
c
t
e
u
r
0
.
3
0
2
0
.
6
9
4
0
.
7
0
2
1
.
0
0
0
0
.
6
7
4
I
n
ter
p
r
etatio
n
:
−
W
ir
e
(
v
11
)
an
d
c
o
n
d
u
cto
r
(
v
22
)
:
h
ig
h
g
lo
b
al
s
co
r
e
(
0
.
5
6
3
an
d
0
.
6
3
6
)
.
T
h
eir
v
er
y
s
tr
o
n
g
s
em
an
t
ic
SB
E
R
T
an
d
J
AC
C
A
R
D
s
im
ilar
i
ty
wit
h
“
cu
r
r
e
n
t
”
.
Sh
o
w
ex
ce
llen
t c
o
h
er
en
ce
with
t
h
e
elec
tr
ical
co
n
tex
t.
−
So
n
(
v
12
)
an
d
d
r
i
v
er
(
v
21
)
:
t
h
eir
s
im
i
lar
ities
ar
e
wea
k
er
an
d
p
r
im
ar
ily
alig
n
ed
with
“
cu
t
”
,
in
d
icatin
g
a
less
r
elev
an
t c
o
h
e
r
en
ce
f
o
r
th
e
elec
tr
ical
co
n
tex
t.
−
Key
f
in
d
in
g
: o
n
ly
wir
e
(
v
11
)
an
d
c
o
n
d
u
ct
o
r
(
v
22
):
d
em
o
n
s
tr
ate
o
p
ti
m
al
alig
n
m
en
t w
ith
th
e
s
en
te
n
ce
.
3.
6
.
Ste
p
4:
s
ens
e
weig
hting
Fo
r
ea
ch
am
b
i
g
u
o
u
s
wo
r
d
s
en
s
e
s
i
,
b
u
ild
a
h
y
b
r
i
d
s
im
ilar
ity
v
ec
to
r
(
T
ab
le
3
)
:
⃗
⃗
⃗
=
0
.
7
×
(
,
ˊ
)
+
0
.
3
×
+
0
.
7
×
(
,
)
0
.
3
×
(
,
)
(
4
)
⃗
⃗
⃗
: g
lo
b
al
v
ec
to
r
r
ep
r
esen
tin
g
th
e
co
n
tex
tu
al
s
im
ilar
ity
f
o
r
th
e
am
b
ig
u
o
u
s
wo
r
d
s
i
.
E
x
am
p
le:
Fo
r
“
co
n
d
u
cte
u
r
”
as
“
d
r
iver
”
:
v
⃗
=
[
0
.
7
×
0
.
843
+
0
.
3
×
0
.
497
,
0
.
7
×
0
.
729
+
0
.
3
×
0
.
371
]
=
[
0
.
739
,
0
.
622
]
T
ab
le
3
.
Vec
to
r
b
ased
co
n
tex
t
u
al
alig
n
m
m
en
t
o
f
wo
r
d
s
en
s
e
Ta
r
g
e
t
w
o
r
d
V
e
c
t
o
r
I
n
t
e
r
p
r
e
t
a
t
i
o
n
F
i
l
s
[
0
.
4
8
6
,
0
.
6
6
6
]
S
t
r
o
n
g
a
l
i
g
n
m
e
n
t
w
i
t
h
e
l
e
c
t
r
i
c
a
l
c
u
r
r
e
n
t
F
i
l
s
[
0
.
5
4
7
,
0
.
3
9
4
]
N
o
me
a
n
i
n
g
f
u
l
c
o
r
r
e
l
a
t
i
o
n
w
i
t
h
e
l
e
c
t
r
i
c
a
l
c
o
n
t
e
x
t
C
o
n
d
u
c
t
e
u
r
[
0
.
7
3
9
,
0
.
6
2
2
]
M
o
d
e
r
a
t
e
t
e
c
h
n
i
c
a
l
r
e
l
e
v
a
n
c
e
C
o
n
d
u
c
t
e
u
r
[
0
.
5
8
3
,
0
.
9
0
8
]
G
o
o
d
c
u
r
r
e
n
t
c
o
r
r
e
l
a
t
i
o
n
T
h
ese
v
ec
to
r
s
allo
w
q
u
an
tita
tiv
e
weig
h
tin
g
o
f
ea
ch
s
en
s
e
b
ased
o
n
its
s
em
an
tic
an
d
lex
ical
co
h
er
en
ce
,
f
ac
ilit
atin
g
th
e
s
elec
tio
n
o
f
th
e
m
o
s
t c
o
n
tex
tu
ally
a
p
p
r
o
p
r
iate
m
ea
n
in
g
.
3
.
7
.
Ste
p
5:
c
ent
ro
id a
na
ly
s
i
s
I
d
en
tify
th
e
s
en
s
e
m
o
s
t
c
o
h
er
en
t
with
th
e
o
v
er
all
c
o
n
tex
t
u
s
in
g
a
g
eo
m
etr
ic
ce
n
ter
.
T
a
b
le
4
s
h
o
ws
C
en
tr
o
id
ca
lcu
latio
n
:
c
o
m
p
u
te
th
e
ar
ith
m
etic
m
ea
n
o
f
th
e
co
o
r
d
in
ates o
f
all
p
o
in
ts
(
all
s
en
s
es c
o
m
b
in
ed
)
.
=
1
∑
=
1
(
5
)
W
h
er
e:
G
=
(
G
x
,
G
y
)
is
th
e
ce
n
tr
o
id
o
r
m
ea
n
ce
n
ter
o
f
th
e
p
o
in
ts
.
V
i
=
(
x
i
,
y
i
)
i
s
th
e
v
ec
to
r
r
e
p
r
esen
tin
g
th
e
i
-
th
p
o
in
t.
n
is
th
e
to
tal
n
u
m
b
er
o
f
p
o
i
n
ts
E
x
am
p
le:
w
ith
4
p
o
in
ts
:
G
=
(
0
.
486
+
0
.
5
4
7
+
0
.
7
3
9
+
0
.
583
4
,
0
.
666
+
0
.
3
94
+
0
.
6
2
2
+
0
.
90
8
4
)
=
(
0
.
589
,
0
.
648
)
E
u
clid
ea
n
d
is
tan
ce
to
t
h
e
ce
n
tr
o
id
:
Fo
r
ea
ch
s
en
s
e
s
i
with
co
o
r
d
in
at
es (
x
i
,
y
i
)
:
(
,
)
=
√
(
−
)
2
+
(
−
)
2
(
6
)
E
x
am
p
le:
d
is
tan
ce
f
o
r
wir
e.
n
.
0
1
(
0
.
4
8
6
,
0
.
6
6
6
)
:
d
=
√
(
0
.
486
−
0
.
589
)
2
+
(
0
.
666
−
0
.
648
)
2
=
√
0
.
0109
=0
.
1
0
4
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
41
,
No
.
3
,
Ma
r
ch
20
2
6
:
9
5
4
-
965
960
T
ab
le
4
.
E
u
clid
e
a
n
d
is
tan
ce
to
th
e
ce
n
tr
o
id
Ta
r
g
e
t
w
o
r
d
sen
s
e
D
i
st
a
n
c
e
I
n
t
e
r
p
r
e
t
a
t
i
o
n
F
i
l
s
v
11
0
.
1
0
4
V
e
r
y
c
l
o
se
t
o
d
e
c
e
n
t
r
o
i
d
F
i
l
s
v
12
0
.
2
5
7
F
a
r
t
h
e
st
f
r
o
m
t
h
e
c
e
n
t
r
o
i
d
C
o
n
d
u
c
t
e
u
r
v
21
0
.
1
5
4
F
u
r
t
h
e
r
f
r
o
m
t
h
e
c
e
n
t
r
o
i
d
C
o
n
d
u
c
t
e
u
r
v
22
0
.
2
6
1
C
l
o
ses
t
t
o
t
h
e
c
e
n
t
r
o
i
d
3
.
8
.
S
t
ep
6:
n
o
rma
liza
t
io
n
T
h
e
p
r
o
b
ab
ilit
y
o
f
a
s
en
s
e
i f
o
r
an
am
b
i
g
u
o
u
s
wo
r
d
is
th
e
r
at
io
o
f
th
e
e
x
p
o
n
en
tial o
f
its
g
lo
b
al
s
co
r
e
to
th
e
s
u
m
o
f
th
e
ex
p
o
n
e
n
tials
o
f
th
e
g
l
o
b
al
s
co
r
es o
f
all
p
o
s
s
ib
le
s
en
s
es j
o
f
th
e
s
am
e
wo
r
d
(
T
ab
le
5
)
.
(
)
=
∑
(
7
)
T
ab
le
5
.
Pro
b
ab
ilit
y
Ta
r
g
e
t
w
o
r
d
G
l
o
b
a
l
sc
o
r
e
P
r
o
b
a
b
i
l
i
t
y
F
i
l
s
0
.
5
6
3
5
3
.
6
%
F
i
l
s
0
.
4
1
7
4
6
.
4
%
C
o
n
d
u
c
t
e
u
r
0
.
6
3
6
5
0
.
3
%
C
o
n
d
u
c
t
e
u
r
0
.
6
2
6
4
9
.
7
%
I
n
ter
p
r
etatio
n
:
t
h
e
So
f
tMa
x
n
o
r
m
aliza
tio
n
t
h
en
co
n
v
er
ts
th
ese
g
lo
b
al
s
co
r
es
in
to
p
r
o
b
ab
ilit
ies,
p
r
o
v
id
i
n
g
a
r
elativ
e
lik
elih
o
o
d
o
f
ea
ch
s
en
s
e
in
th
e
g
iv
e
n
co
n
tex
t.
I
n
th
e
tab
le
ab
o
v
e,
th
e
s
e
n
s
e
wir
e
r
ec
eiv
es a
p
r
o
b
a
b
ilit
y
o
f
5
3
.
6
%,
wh
ile
th
e
s
en
s
e
s
o
n
h
as
4
6
.
4
%.
T
h
is
in
d
icate
s
th
at
wir
e
is
s
lig
h
tly
m
o
r
e
lik
ely
in
th
e
co
n
tex
t,
b
u
t
b
o
t
h
s
en
s
es
ar
e
s
till
p
lau
s
ib
le.
Pro
b
ab
ilit
ies
f
o
r
o
th
er
a
m
b
ig
u
o
u
s
wo
r
d
s
a
r
e
in
ter
p
r
eted
s
im
ilar
ly
,
g
iv
in
g
a
n
o
r
m
alize
d
,
co
m
p
ar
at
iv
e
m
ea
s
u
r
e
o
f
s
en
s
e
r
elev
an
c
e.
3
.
9
.
Ste
p
7:
c
o
m
pu
t
a
t
io
n
o
f
t
he
g
lo
ba
l
J
-
f
uzzy
weig
hting
s
co
re
a
nd
s
em
a
ntic
s
im
ila
rit
y
m
ea
s
urem
en
t
bet
wee
n lex
ica
l
y
s
im
ila
r
s
ent
ence
s
T
h
e
g
lo
b
al
J
-
f
u
zz
y
weig
h
ted
s
co
r
e
co
m
b
in
es f
u
zz
y
lex
ical
s
im
ilar
ity
an
d
s
en
s
e
p
r
o
b
ab
ilit
y
f
o
r
W
SD
.
C
an
d
id
ate
d
ef
in
itio
n
s
ar
e
s
co
r
ed
u
s
in
g
f
u
zz
y
J
ac
ca
r
d
a
n
d
SB
E
R
T
s
im
ilar
ity
,
n
o
r
m
alize
d
v
ia
So
f
tMa
x
,
an
d
th
e
f
in
al
s
co
r
e
is
th
e
weig
h
ted
a
v
e
r
ag
e
o
f
f
u
zz
y
J
ac
ca
r
d
v
alu
es
m
u
ltip
lied
b
y
s
en
s
e
p
r
o
b
a
b
ilit
ies
J
F
uzz
y
W
ei
g
ht
ed
=
1
N
∑
P
(
s
i
)
×
J
a
c
c
a
r
d
(
s
i
)
=
1
(
8
)
Her
e,
d
en
o
tes
th
e
s
en
s
e
p
r
o
b
a
b
ilit
y
an
d
r
ep
r
esen
ts
th
e
f
u
zz
y
lex
ical
s
im
ilar
ity
b
etwe
en
c
o
n
tex
t
an
d
th
e
s
en
s
e
d
ef
in
itio
n
.
T
h
is
m
ea
s
u
r
e
ca
p
tu
r
es
s
em
an
tic
s
i
m
ilar
ity
ev
en
f
o
r
lex
ically
clo
s
e
s
en
ten
ce
s
;
f
o
r
in
s
tan
ce
,
“
L
es
f
ils
d
u
c
o
n
d
u
c
teu
r
o
n
t
co
u
p
é
le
C
o
u
r
a
nt
”
d
o
es
n
o
t
r
ea
c
h
f
u
ll
s
im
ilar
ity
b
ec
au
s
e
am
b
ig
u
o
u
s
wo
r
d
s
en
s
es
(
f
ils
:
s
o
n
v
s
f
ils
:
wir
es)
af
f
ec
t
th
e
s
co
r
e.
T
h
u
s
,
th
e
g
lo
b
al
J
-
f
u
zz
y
weig
h
ted
s
co
r
e
r
ef
lects
b
o
th
lex
ical
o
v
er
lap
an
d
s
en
s
e
p
la
u
s
ib
ilit
y
,
in
teg
r
atin
g
co
n
tex
t
u
al
an
d
s
em
an
tic
u
n
d
er
s
tan
d
in
g
,
e
n
ab
lin
g
d
is
tin
ctio
n
o
f
lex
ically
id
en
tical
b
u
t
s
e
m
an
tically
d
if
f
er
e
n
t
s
en
ten
ce
s
,
an
d
o
f
f
e
r
in
g
a
m
o
r
e
h
u
m
an
-
lik
e
an
d
r
o
b
u
s
t
m
ea
s
u
r
e
th
an
p
u
r
ely
le
x
ical
ap
p
r
o
ac
h
es.
3
.
10
.
Ste
p
8:
v
a
ria
nt
pro
j
ec
t
io
n
T
o
r
ep
r
esen
t
t
h
e
p
o
te
n
tial
m
e
an
in
g
s
o
f
a
m
b
ig
u
o
u
s
ter
m
s
li
k
e
f
ils
an
d
co
n
d
u
cte
u
r
,
a
v
ec
t
o
r
s
p
ac
e
is
b
u
ilt
with
two
co
n
tex
tu
al
d
im
en
s
io
n
s
:
th
e
x
-
ax
is
en
co
d
es
weig
h
ted
f
u
zz
y
J
ac
ca
r
d
a
n
d
SB
E
R
T
s
im
ilar
ity
w
ith
co
u
p
é
(
elec
tr
ical
in
te
r
r
u
p
tio
n
)
,
an
d
th
e
y
-
ax
is
en
c
o
d
es th
e
s
a
m
e
s
im
ilar
ity
with
co
u
r
an
t
(
elec
tr
ical
f
lo
w)
.
I
n
ter
p
r
etatio
n
:
T
h
is
d
u
al
-
m
etr
ic
ap
p
r
o
ac
h
c
o
m
b
in
es
f
u
zz
y
J
ac
ca
r
d
an
d
SB
E
R
T
to
p
o
s
itio
n
ea
ch
s
en
s
e
in
th
e
co
n
tex
tu
al
v
ec
to
r
s
p
ac
e.
C
en
tr
o
id
an
aly
s
is
s
h
o
ws
th
at
co
n
d
u
cteu
r
(
élec
tr
iq
u
e)
a
n
d
f
ils
(
élec
tr
iq
u
e)
ar
e
clo
s
est
to
th
e
av
er
a
g
e
co
n
tex
t,
c
o
n
f
ir
m
in
g
th
eir
tec
h
n
ical
r
elev
a
n
ce
,
wh
ile
f
ils
(
f
am
ilial)
an
d
co
n
d
u
cteu
r
(
ch
au
f
f
eu
r
)
lie
f
ar
th
er
awa
y
,
ex
clu
d
in
g
th
em
f
r
o
m
th
e
elec
tr
ical
d
o
m
ain
.
T
h
is
v
ec
to
r
-
b
ased
an
aly
s
is
v
alid
ates
o
p
tim
al
d
is
am
b
ig
u
atio
n
f
o
r
th
e
co
n
tex
t
.
Geo
m
etr
ic
an
aly
s
is
o
f
c
o
n
tex
tu
al
wo
r
d
s
en
s
es
as
s
h
o
wn
in
Fig
u
r
e
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mea
s
u
r
in
g
s
ema
n
tic
s
imila
r
ity
in
lexi
ca
lly
id
en
tica
l
…
(
B
tis
s
a
m
E
l J
a
n
a
ti
)
961
Fig
u
r
e
2
.
Geo
m
etr
ic
an
aly
s
is
o
f
co
n
te
x
tu
al
wo
r
d
s
en
s
es
3
.
1
1
.
Co
m
pu
t
a
t
i
o
na
l c
o
m
plex
it
y
a
nd
s
ca
la
bil
it
y
T
h
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
v
ar
i
es
ac
r
o
s
s
its
co
m
p
o
n
en
ts
.
T
o
k
en
izatio
n
an
d
lem
m
atiza
tio
n
s
ca
le
lin
ea
r
ly
with
s
en
ten
ce
len
g
th
,
wh
ile
lex
ico
n
lo
o
k
u
p
d
ep
e
n
d
s
o
n
th
e
n
u
m
b
e
r
o
f
am
b
ig
u
o
u
s
to
k
en
s
.
T
h
e
weig
h
ted
f
u
zz
y
J
ac
ca
r
d
co
m
p
u
tatio
n
s
ca
l
es
with
th
e
s
ize
o
f
g
lo
s
s
es
an
d
co
n
tex
tu
a
l
to
k
en
s
,
an
d
SB
E
R
T
in
f
er
en
ce
d
ep
en
d
s
o
n
b
o
t
h
s
en
ten
ce
len
g
th
an
d
em
b
e
d
d
in
g
d
im
e
n
s
io
n
ality
.
C
en
tr
o
id
an
d
d
is
tan
ce
ca
lcu
latio
n
s
s
ca
le
with
th
e
n
u
m
b
er
o
f
ca
n
d
id
ate
s
en
s
es,
an
d
th
e
So
f
tMa
x
lay
e
r
with
th
e
n
u
m
b
er
o
f
s
en
s
e
p
r
o
b
ab
ilit
ies to
co
m
p
u
te
.
Fro
m
a
p
r
ac
tical
s
tan
d
p
o
in
t,
ef
f
icien
cy
ca
n
b
e
im
p
r
o
v
ed
b
y
ca
ch
in
g
em
b
ed
d
in
g
s
an
d
in
ter
m
ed
iate
r
esu
lts
,
p
er
f
o
r
m
in
g
b
atch
e
d
SB
E
R
T
in
f
er
en
ce
o
n
GPU,
an
d
p
ar
allelizin
g
co
m
p
u
tatio
n
s
ac
r
o
s
s
to
k
en
s
o
r
s
en
ten
ce
s
.
T
h
ese
o
p
tim
izatio
n
s
en
s
u
r
e
s
ca
lab
ilit
y
f
o
r
lar
g
e
r
d
atasets
an
d
lo
n
g
er
te
x
tu
al
in
p
u
ts
.
3
.
1
2
.
I
nte
rpre
t
a
bil
it
y
Ou
r
ap
p
r
o
ac
h
p
r
o
v
id
es in
ter
p
r
etab
ilit
y
at
th
r
ee
co
m
p
lem
en
ta
r
y
lev
els:
−
L
ex
ical
lev
el:
t
h
e
weig
h
ted
f
u
zz
y
J
ac
ca
r
d
id
en
tifie
s
k
ey
t
o
k
en
s
in
W
o
r
d
Net
d
ef
in
itio
n
s
,
r
ev
ea
lin
g
th
eir
r
elativ
e
im
p
o
r
tan
ce
in
s
im
ilar
ity
ca
lcu
latio
n
.
−
Vec
to
r
an
d
d
is
tan
ce
lev
el:
s
e
n
ten
ce
s
ar
e
r
e
p
r
esen
ted
as
h
y
b
r
id
v
ec
to
r
s
co
m
b
in
in
g
lex
ic
al
an
d
s
em
an
tic
f
ea
tu
r
es.
T
h
e
ce
n
tr
o
id
s
er
v
es
as
a
s
em
an
tic
r
ef
er
en
ce
p
o
in
t,
wh
er
e
s
h
o
r
ter
E
u
c
lid
ea
n
d
is
tan
ce
s
in
d
icate
h
ig
h
er
s
im
ilar
ity
an
d
lo
n
g
er
d
i
s
tan
ce
s
s
h
o
w
s
em
an
tic
d
iv
er
g
en
ce
.
−
Pro
b
ab
ilis
tic
lev
el:
t
h
e
So
f
t
Ma
x
f
u
n
ctio
n
co
n
v
er
ts
r
aw
s
im
ilar
ity
s
co
r
es
in
to
p
r
o
b
ab
il
ity
d
is
tr
ib
u
tio
n
s
,
en
ab
lin
g
clea
r
c
o
m
p
a
r
is
o
n
b
et
wee
n
s
y
n
s
ets an
d
p
r
o
v
id
in
g
a
p
r
o
b
a
b
ilis
tic
b
asis
f
o
r
co
m
p
o
n
en
t f
u
s
io
n
.
4.
E
XP
E
R
I
M
E
N
T
A
L
VAL
I
D
AT
I
O
N
4
.
1
.
H
u
m
a
n e
v
a
lua
t
io
n f
ra
mewo
rk
A
r
ig
o
r
o
u
s
ev
alu
atio
n
p
r
o
to
c
o
l
was
im
p
lem
en
ted
u
s
in
g
a
0
-
1
0
0
s
ca
le
with
n
ativ
e
Fre
n
ch
s
p
ea
k
er
s
,
in
clu
d
in
g
lin
g
u
is
tics
an
d
NL
P
s
p
ec
iali
s
ts
.
T
h
e
m
eth
o
d
o
lo
g
y
d
em
o
n
s
tr
ated
r
o
b
u
s
t
r
eliab
ilit
y
with
Flei
s
s
’
Kap
p
a
o
f
0
.
4
8
(
m
o
d
er
ate
ag
r
e
em
en
t)
an
d
test
-
r
etest
co
r
r
elatio
n
o
f
0
.
8
2
,
s
u
p
p
o
r
ted
b
y
co
m
p
r
eh
en
s
iv
e
q
u
ality
co
n
tr
o
l m
ea
s
u
r
es to
en
s
u
r
e
d
ata
v
alid
ity
.
4
.
2
.
Co
m
pa
ra
t
iv
e
s
y
s
t
m
e
s
p
ec
if
ica
t
io
n
T
h
e
ex
p
er
i
m
e
n
tal
ev
alu
atio
n
i
n
co
r
p
o
r
ated
co
m
p
r
eh
e
n
s
iv
e
s
tate
-
of
-
th
e
-
a
r
t
s
y
s
tem
s
in
clu
d
in
g
GPT
-
4
(
g
p
t
-
4
-
0
6
1
3
)
,
Gem
in
i
Pro
(
g
em
in
i
-
1
.
5
-
p
r
o
)
,
Per
p
le
x
ity
AI
(
2
0
2
4
-
0
8
API
)
,
Dee
p
Seek
C
h
at
(
d
ee
p
s
ee
k
-
llm
-
67b
-
c
h
at)
,
alo
n
g
with
r
esear
ch
s
y
s
tem
s
Sim
C
SE
-
f
r
an
d
T
5
-
E
n
co
d
er
.
All
s
y
s
tem
s
u
n
d
er
wen
t
id
en
tical
p
r
ep
r
o
ce
s
s
in
g
p
ip
elin
es
with
u
n
if
ie
d
7
6
8
-
d
im
en
s
io
n
al
em
b
ed
d
in
g
s
,
u
s
in
g
co
s
in
e
s
im
ilar
ity
with
L
2
n
o
r
m
aliza
tio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
41
,
No
.
3
,
Ma
r
ch
20
2
6
:
9
5
4
-
965
962
4
.
3
.
E
x
perim
ent
a
l r
esu
lt
s
A
co
m
p
r
eh
en
s
iv
e
an
aly
s
is
o
f
3
3
s
en
te
n
ce
s
co
m
p
ar
ed
h
u
m
an
ev
alu
atio
n
s
with
AI
-
g
en
er
ated
s
im
ilar
ity
s
co
r
es a
n
d
o
u
r
h
y
b
r
i
d
ap
p
r
o
ac
h
,
as d
etailed
in
T
ab
l
e
6
.
T
ab
le
6
.
E
v
alu
atio
n
h
u
m
an
ju
d
g
m
en
ts
v
s
ar
tific
ial
in
tellig
en
ce
(
AI
)
s
y
s
te
m
s
v
s
h
y
b
r
i
d
ap
p
r
o
a
ch
S
e
n
t
e
n
c
e
s
A
p
p
r
o
a
c
h
H
u
ma
n
AI
H
u
ma
n
-
AI
H
u
ma
n
-
a
p
p
r
o
a
c
h
L
e
s f
i
l
s
d
u
c
o
n
d
u
c
t
e
u
r
o
n
t
c
o
u
p
é
l
e
c
o
u
ra
n
t
0
.
7
2
0
.
3
2
1
.
0
0
0
.
7
9
0
.
4
3
L
e
m
é
d
e
c
i
n
a
d
e
m
a
n
d
é
u
n
e
ra
d
i
o
d
u
b
ra
s
0
.
7
3
1
.
0
0
1
.
0
0
0
.
0
0
0
.
6
1
L
a
c
a
r
t
e
d
u
r
o
i
e
st
p
e
r
d
u
e
0
.
3
7
0
.
0
4
1
.
0
0
0
.
6
2
0
.
7
2
L
a
p
o
l
i
c
e
e
st
v
u
e
a
v
e
c
u
n
e
b
a
rr
e
0
.
2
4
0
.
6
1
1
.
0
0
0
.
3
9
0
.
6
2
L
e
res
p
o
n
s
a
b
l
e
d
e
n
o
t
re
p
o
s
t
e
a
v
i
s
i
t
é
l
e
p
a
rc
0
.
5
6
0
.
8
1
1
.
0
0
0
.
1
8
0
.
8
2
I
l
f
a
u
t
a
j
o
u
t
e
r
p
l
u
si
e
u
rs
ser
v
e
u
rs
p
o
u
r
ren
f
o
rce
r
n
o
t
re
syst
è
m
e
0
.
7
3
0
.
5
6
1
.
0
0
0
.
4
4
0
.
5
9
L
e
p
r
o
f
e
sse
u
r
a
c
a
ssé
l
a
rè
g
l
e
e
n
c
l
a
s
se
0
.
7
9
0
.
0
0
1
.
0
0
1
.
0
0
0
.
0
0
Ap
rès
l
e
v
o
l
d
e
l
a
n
u
i
t
,
l
e
s
b
a
l
l
e
s s
o
n
t
l
a
n
c
é
e
s
a
u
c
i
e
l
0
.
9
8
0
.
6
4
1
.
0
0
0
.
3
7
0
.
2
3
L
e
s f
e
u
i
l
l
e
s
so
n
t
sa
u
v
e
g
a
r
d
é
e
s s
u
r
l
a
t
a
b
l
e
0
.
3
7
0
.
8
6
1
.
0
0
0
.
1
2
0
.
8
3
L
a
t
a
b
l
e
e
s
t
t
rès
u
t
i
l
e
p
o
u
r
l
e
t
r
a
v
a
i
l
a
u
b
u
r
e
a
u
0
.
8
7
0
.
7
7
1
.
0
0
0
.
2
3
0
.
3
4
L
e
t
o
u
r
d
u
ro
i
e
s
t
d
a
n
g
e
r
e
u
x
0
.
1
2
0
.
4
9
1
.
0
0
0
.
5
1
0
.
6
0
L
a
f
e
n
ê
t
r
e
e
st
r
e
st
é
e
o
u
v
e
rt
e
p
e
n
d
a
n
t
q
u
’
i
l
m
a
rc
h
e
0
.
8
6
0
.
7
6
1
.
0
0
0
.
2
4
0
.
3
3
L
e
s p
u
c
e
s s
o
n
t
v
i
s
i
b
l
e
s
s
u
r l
a
c
a
r
t
e
0
.
0
0
0
.
2
8
1
.
0
0
0
.
8
3
0
.
8
3
O
n
a
d
é
c
o
u
v
e
rt
l
a
v
a
l
e
u
r
d
e
n
o
t
re
t
o
u
r a
u
c
o
u
rs
d
e
l
’
h
i
s
t
o
i
re
0
.
5
9
0
.
5
7
1
.
0
0
0
.
3
8
0
.
5
9
L
a
s
o
m
m
e
d
e
s
c
h
a
r
g
e
s
e
s
t
l
o
u
r
d
e
0
.
1
7
0
.
3
2
1
.
0
0
0
.
6
8
0
.
4
1
L
a
r
o
u
e
a
r
o
u
l
é
p
r
è
s
d
u
t
r
a
i
n
0
.
9
7
0
.
6
5
1
.
0
0
0
.
2
5
0
.
0
5
S
a
v
o
i
x
c
o
m
p
t
e
d
a
n
s
l
a
d
é
c
i
si
o
n
f
i
n
a
l
e
0
.
8
3
0
.
5
6
1
.
0
0
0
.
4
5
0
.
2
5
I
l
a
p
o
sé
u
n
e
c
o
u
c
h
e
s
u
r
l
a
t
a
b
l
e
0
.
6
5
0
.
4
0
1
.
0
0
0
.
6
1
0
.
3
1
L
e
p
l
a
n
d
u
s
p
e
c
t
a
c
l
e
a
é
t
é
i
m
p
r
i
m
é
s
u
r l
a
t
a
b
l
e
0
.
7
4
0
.
5
8
1
.
0
0
0
.
4
3
0
.
2
4
I
l
a
a
c
h
e
t
é
u
n
e
l
i
v
r
e
a
u
m
a
r
c
h
é
a
v
e
c
s
e
s f
i
l
s
0
.
4
0
0
.
7
9
1
.
0
0
0
.
2
1
0
.
4
9
L
a
m
i
n
e
e
s
t
u
sée
a
p
r
è
s
b
e
a
u
c
o
u
p
d
’
u
s
a
g
e
0
.
4
3
0
.
6
6
1
.
0
0
0
.
3
3
0
.
4
4
L
a
p
ress
i
o
n
s
u
r
l
e
c
o
u
rs
e
s
t
i
n
q
u
i
é
t
a
n
t
e
0
.
1
7
0
.
5
6
1
.
0
0
0
.
4
5
0
.
9
2
L
e
s e
n
f
a
n
t
s
p
o
rt
e
n
t
u
n
e
o
r
a
n
g
e
à
l
a
m
a
i
t
ress
e
p
rès
d
u
p
a
r
c
0
.
4
2
0
.
7
5
1
.
0
0
0
.
3
3
0
.
3
2
El
l
e
a
é
c
ri
t
l
a
l
e
t
t
r
e
e
n
c
a
p
i
t
a
l
e
c
e
m
a
t
i
n
0
.
4
9
0
.
3
4
1
.
0
0
0
.
6
7
0
.
3
1
I
l
a
p
e
rd
u
l
a
p
i
l
e
a
u
c
o
u
rs
d
u
m
a
t
c
h
0
.
4
5
0
.
3
1
1
.
0
0
0
.
7
0
1
.
0
0
L
a
m
a
n
c
h
e
a
p
ri
s
p
r
è
s
d
u
f
e
u
1
.
0
0
0
.
6
9
1
.
0
0
0
.
3
1
0
.
1
0
L
a
b
ra
n
c
h
e
q
u
i
m
è
n
e
a
u
d
r
o
i
t
e
s
t
b
i
e
n
d
é
f
i
n
i
e
0
.
6
4
0
.
2
1
1
.
0
0
0
.
6
9
0
.
3
9
L
e
v
o
i
l
e
e
st
t
o
m
b
é
a
u
m
o
m
e
n
t
d
u
sp
e
c
t
a
c
l
e
0
.
4
3
0
.
4
1
1
.
0
0
0
.
6
0
0
.
2
1
J
a
v
a
p
o
ssè
d
e
p
l
u
si
e
u
rs
p
o
rt
s
a
c
t
i
f
s
à
g
é
r
e
r
0
.
4
2
0
.
5
7
1
.
0
0
0
.
4
4
0
.
4
3
D
u
ra
n
t
l
a
s
a
i
s
o
n
d
e
réc
o
l
t
e
,
l
a
p
ê
c
h
e
a
c
a
ss
é
l
a
b
r
a
n
c
h
e
0
.
4
4
0
.
5
8
1
.
0
0
0
.
2
5
0
.
4
1
Statis
t
ical
an
aly
s
is
s
h
o
ws
th
at
th
e
d
if
f
er
en
ce
b
etwe
en
AI
s
y
s
tem
s
an
d
h
u
m
an
ev
alu
atio
n
s
is
h
ig
h
ly
s
ig
n
if
ican
t
(
p
=
0
.
0
0
0
0
5
9
<
0
.
0
0
1
)
.
T
h
u
s
,
th
e
n
u
ll
h
y
p
o
th
esis
(
H₀)
is
r
ejec
ted
,
in
d
icatin
g
th
at
cu
r
r
en
t
s
y
s
tem
s
s
ig
n
if
ican
tly
d
iv
er
g
e
f
r
o
m
h
u
m
an
j
u
d
g
m
e
n
ts
,
wh
ile
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
ap
p
r
o
ac
h
ac
h
iev
es
b
etter
co
n
v
er
g
en
ce
.
T
h
e
ef
f
ec
t
s
ize
(
C
o
h
en
’
s
d
=
-
0
.
8
0
)
c
o
n
f
ir
m
s
a
s
tr
o
n
g
e
f
f
ec
t,
d
em
o
n
s
tr
atin
g
th
at
th
e
im
p
r
o
v
em
e
n
t is b
o
th
s
tatis
tical
ly
an
d
p
r
ac
tically
m
ea
n
in
g
f
u
l.
C
o
n
f
id
en
ce
in
ter
v
als f
u
r
th
er
s
u
p
p
o
r
t th
e
r
o
b
u
s
tn
ess
o
f
th
ese
f
in
d
in
g
s
:
−
MA
E
(
Hy
b
r
id
a
p
p
r
o
ac
h
)
: 4
0
.
0
2
[
3
9
.
0
5
–
4
0
.
9
9
]
.
−
MA
E
(
R
ef
er
en
ce
AI
)
: 4
5
.
3
2
[
4
3
.
3
9
–
4
7
.
2
6
]
.
T
h
e
av
er
ag
e
r
ed
u
ctio
n
o
f
5
.
3
1
p
o
in
ts
co
r
r
esp
o
n
d
s
to
a
r
el
ativ
e
im
p
r
o
v
em
e
n
t
o
f
1
1
.
7
%.
Sin
ce
th
e
co
n
f
id
en
ce
in
ter
v
al
[
-
7
.
6
4
,
-
2
.
9
7
]
ex
clu
d
es
ze
r
o
,
th
e
o
b
s
er
v
ed
e
f
f
ec
t
is
r
o
b
u
s
t
a
n
d
r
eliab
le.
Mo
r
eo
v
er
,
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
ac
h
iev
es
th
e
lo
west
MA
E
an
d
r
o
o
t
m
ea
n
s
q
u
ar
ed
e
r
r
o
r
(
R
MSE
)
,
th
e
h
ig
h
est
s
em
an
tic
ac
cu
r
ac
y
(
0
.
8
7
5
)
,
a
n
d
th
e
b
es
t
m
ea
n
r
an
k
(
1
.
8
)
,
c
o
n
f
ir
m
in
g
its
s
tab
ilit
y
an
d
clo
s
en
ess
to
h
u
m
an
r
ea
s
o
n
in
g
.
T
ab
le
7
co
m
p
ar
es
th
e
h
y
b
r
id
ap
p
r
o
ac
h
with
SimCS
E
an
d
T
5
ac
r
o
s
s
m
u
ltip
le
m
etr
ics,
ass
es
s
in
g
s
em
an
tic
alig
n
m
en
t,
er
r
o
r
m
i
n
im
izatio
n
,
an
d
co
r
r
elatio
n
with
h
u
m
a
n
ju
d
g
m
e
n
ts
:
W
h
ile
T
5
ac
h
iev
es
h
ig
h
e
r
c
o
r
r
elatio
n
co
ef
f
icien
ts
d
u
e
to
its
lar
g
e
-
s
ca
le
p
r
etr
ain
in
g
,
t
h
e
h
y
b
r
id
ap
p
r
o
ac
h
o
u
tp
er
f
o
r
m
s
b
o
th
m
o
d
els
in
MA
E
,
R
MSE
,
an
d
s
em
an
tic
ac
cu
r
ac
y
,
co
n
f
ir
m
in
g
i
ts
g
r
ea
ter
s
tab
ilit
y
an
d
in
ter
p
r
etab
ilit
y
.
T
h
e
lo
we
r
er
r
o
r
m
ar
g
in
s
i
n
d
icate
th
at
i
t
n
o
t
o
n
ly
ca
p
tu
r
es
g
en
e
r
al
s
e
m
an
tic
tr
en
d
s
b
u
t
also
f
in
e
-
g
r
ain
e
d
d
is
tin
ctio
n
s
i
n
m
ea
n
in
g
.
T
ab
le
7
.
C
o
m
p
a
r
ativ
e
p
er
f
o
r
m
an
ce
-
h
y
b
r
id
a
p
p
r
o
ac
h
v
s
.
m
o
d
er
n
m
eth
o
d
s
M
e
t
r
i
c
H
y
b
r
i
d
a
p
p
r
o
a
c
h
S
i
mCSE
T5
B
e
st
S
p
e
a
r
m
a
n
c
o
r
r
e
l
a
t
i
o
n
0
.
7
6
2
0
.
7
1
4
0
.
9
0
5
T5
P
e
a
r
so
n
c
o
r
r
e
l
a
t
i
o
n
0
.
8
4
1
0
.
7
0
1
0
.
9
2
6
T5
M
A
E
0
.
1
3
3
0
.
3
0
9
0
.
3
7
6
H
y
b
r
i
d
R
M
S
E
0
.
2
1
1
0
.
4
0
8
0
.
4
6
5
H
y
b
r
i
d
S
e
ma
n
t
i
c
a
c
c
u
r
a
c
y
0
.
8
7
5
0
.
5
0
0
0
.
5
0
0
H
y
b
r
i
d
M
e
a
n
r
a
n
k
1
.
8
2
.
8
2
.
4
H
y
b
r
i
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mea
s
u
r
in
g
s
ema
n
tic
s
imila
r
ity
in
lexi
ca
lly
id
en
tica
l
…
(
B
tis
s
a
m
E
l J
a
n
a
ti
)
963
5.
DIS
CU
S
SI
O
N
Ou
r
h
y
b
r
id
ap
p
r
o
ac
h
d
e
m
o
n
s
tr
ates
s
u
p
er
io
r
s
em
an
tic
co
h
er
en
ce
with
h
u
m
an
ju
d
g
m
e
n
ts
th
r
o
u
g
h
s
tr
o
n
g
c
o
r
r
elatio
n
s
(
Sp
ea
r
m
a
n
:
0
.
7
6
2
,
Pear
s
o
n
:
0
.
8
4
1
)
an
d
s
ig
n
if
ican
tly
r
e
d
u
ce
d
er
r
o
r
s
(
MA
E
:
0
.
1
3
3
v
s
0
.
3
0
9
-
0
.
3
7
6
f
o
r
b
aselin
es),
ac
h
iev
ed
v
ia
ad
a
p
tiv
e
in
teg
r
atio
n
o
f
s
y
m
b
o
lic
an
d
n
eu
r
al
co
m
p
o
n
en
ts
th
at
d
y
n
am
ically
r
esp
o
n
d
s
to
am
b
ig
u
ity
p
atter
n
s
.
W
h
ile
T
5
a
ch
iev
es
h
ig
h
er
co
r
r
elatio
n
s
(
0
.
9
0
5
Sp
ea
r
m
an
)
,
it
s
y
s
tem
atica
lly
o
v
er
esti
m
ates
lex
ical
s
im
ilar
ity
,
wh
er
ea
s
o
u
r
m
eth
o
d
m
ain
tain
s
b
o
th
co
n
tex
tu
al
s
en
s
itiv
ity
an
d
cr
u
cial
in
ter
p
r
eta
b
ilit
y
,
d
esp
i
te
f
ac
in
g
lim
itat
io
n
s
in
cr
o
s
s
-
lin
g
u
is
tic
ad
ap
tatio
n
an
d
h
an
d
lin
g
co
m
p
le
x
p
r
ag
m
atic
am
b
i
g
u
ities
.
T
h
ese
f
in
d
in
g
s
ch
allen
g
e
co
n
v
en
t
io
n
al
s
ca
lin
g
p
ar
ad
i
g
m
s
b
y
d
em
o
n
s
tr
atin
g
th
at
s
o
p
h
is
ticated
ar
ch
itectu
r
al
d
e
s
ig
n
ca
n
ac
h
iev
e
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
co
m
p
ar
ed
to
s
im
p
l
y
in
cr
ea
s
in
g
m
o
d
el
p
ar
am
eter
s
,
th
er
e
b
y
en
a
b
lin
g
m
o
r
e
r
eso
u
r
ce
-
ef
f
icien
t
d
ep
lo
y
m
en
t.
6.
CO
NCLU
SI
O
N
T
h
is
s
tu
d
y
p
r
esen
ts
a
h
y
b
r
id
ap
p
r
o
ac
h
f
o
r
Fre
n
ch
s
em
a
n
tic
s
im
ilar
ity
an
d
W
SD
,
co
m
b
in
in
g
f
u
zz
y
J
ac
ca
r
d
lex
ical
m
atch
in
g
wit
h
SB
E
R
T
em
b
ed
d
in
g
s
.
E
x
p
e
r
im
en
tal
v
alid
atio
n
o
n
3
3
am
b
ig
u
o
u
s
s
en
ten
ce
s
,
ev
alu
ated
b
y
n
ativ
e
s
p
ea
k
e
r
s
,
d
em
o
n
s
tr
ated
th
at
th
e
h
y
b
r
id
m
o
d
el
s
ig
n
if
ican
tly
o
u
tp
e
r
f
o
r
m
s
ex
is
tin
g
AI
s
y
s
tem
s
,
ac
h
iev
in
g
an
1
1
.
7
% r
ed
u
ctio
n
in
MA
E
co
m
p
ar
ed
t
o
C
h
atGPT
an
d
Gem
in
i (
t =
-
4
.
6
2
,
p
=
5
.
9
2
×
1
0
⁻⁵)
.
T
h
e
ap
p
r
o
ac
h
p
r
o
d
u
ce
s
in
ter
p
r
etab
le,
co
n
tex
t
-
s
en
s
itiv
e
r
esu
lts
,
p
ar
ticu
lar
ly
f
o
r
lex
ic
ally
id
en
tical
b
u
t
s
em
an
tically
d
iv
er
g
e
n
t
s
en
ten
ce
s
,
d
em
o
n
s
tr
atin
g
s
tr
o
n
g
ali
g
n
m
en
t
with
h
u
m
an
ju
d
g
m
en
ts
an
d
tr
a
n
s
p
ar
en
t
d
ec
is
io
n
-
m
ak
in
g
.
Ou
r
ap
p
r
o
ac
h
ch
an
g
es
th
e
s
tate
o
f
th
e
ar
t
b
y
in
tr
o
d
u
ci
n
g
d
y
n
am
ic
in
teg
r
atio
n
b
etwe
en
s
y
m
b
o
lic
an
d
n
eu
r
al
m
eth
o
d
s
,
wh
er
ea
s
p
r
e
v
io
u
s
wo
r
k
f
av
o
r
ed
s
tatic
co
m
b
in
atio
n
s
o
r
p
u
r
el
y
n
eu
r
al
ar
ch
itectu
r
es.
T
h
is
in
n
o
v
atio
n
en
ab
les
co
n
tex
tu
al
am
b
ig
u
ity
r
eso
lu
tio
n
th
at
s
u
r
p
ass
es
th
e
lim
itatio
n
s
o
f
co
n
v
en
tio
n
al
ap
p
r
o
ac
h
es,
wh
ile
m
ain
tain
in
g
th
e
i
n
ter
p
r
e
tab
ilit
y
o
f
ten
s
ac
r
if
iced
in
lar
g
e
-
s
ca
le
m
o
d
els.
T
h
e
h
y
b
r
id
f
r
am
ewo
r
k
co
n
tr
i
b
u
tes
th
eo
r
etica
lly
b
y
b
r
id
g
i
n
g
s
y
m
b
o
lic
lin
g
u
is
tic
k
n
o
wled
g
e
with
n
eu
r
al
c
o
n
tex
tu
al
m
o
d
elin
g
,
a
n
d
p
r
ac
tically
b
y
im
p
r
o
v
in
g
NL
P
task
s
s
u
ch
as
m
ac
h
in
e
t
r
an
s
latio
n
,
q
u
esti
o
n
an
s
wer
in
g
,
an
d
s
em
an
tic
s
ea
r
ch
with
ac
cu
r
ate
an
d
ex
p
lain
ab
le
s
en
ten
ce
-
lev
el
r
ep
r
esen
t
atio
n
s
.
L
im
itatio
n
s
in
clu
d
e
th
e
r
estricte
d
e
v
alu
at
io
n
s
et
an
d
f
o
cu
s
o
n
Fre
n
c
h
,
wh
ile
f
u
tu
r
e
p
er
s
p
ec
tiv
es
in
v
o
lv
e
m
u
ltil
in
g
u
al
ad
ap
tatio
n
,
s
ca
lin
g
to
lar
g
e
r
b
en
ch
m
ar
k
s
,
an
d
in
teg
r
atio
n
o
f
ex
p
lain
ab
le
m
o
d
u
les
to
en
h
a
n
ce
in
ter
p
r
etab
ilit
y
an
d
u
s
er
tr
u
s
t.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
was su
p
p
o
r
ted
b
y
I
b
n
T
o
f
ail
Un
iv
e
r
s
ity
as p
ar
t
o
f
th
e
d
o
cto
r
al
r
esear
ch
p
r
o
je
ct.
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
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
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
B
tis
s
am
E
l
J
an
ati
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
A
d
il
E
n
aa
n
a
i
✓
✓
✓
✓
✓
Fad
o
u
a
Gh
an
im
i
✓
✓
✓
✓
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
ST
A
T
E
M
E
NT
T
h
e
au
th
o
r
s
d
ec
lar
e
th
at
th
ey
h
av
e
n
o
k
n
o
w
n
co
m
p
etin
g
f
in
an
cial
in
ter
ests
o
r
p
er
s
o
n
al
r
el
atio
n
s
h
ip
s
th
at
co
u
ld
h
av
e
ap
p
ea
r
ed
t
o
in
f
lu
en
ce
th
e
wo
r
k
r
e
p
o
r
te
d
in
t
h
is
p
ap
er
.
Evaluation Warning : The document was created with Spire.PDF for Python.