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.
3
8
,
No
.
1
,
A
p
r
il
20
2
5
,
p
p
.
439
~
446
I
SS
N:
2
502
-
4
7
52
,
DOI
: 1
0
.
1
1
5
9
1
/ijee
cs
.v
3
8
.
i
1
.
pp
439
-
4
4
6
439
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs
.
ia
esco
r
e.
co
m
Perf
o
rma
nce
ana
ly
sis
of dif
fe
rent
B
ERT
implemen
ta
tion for
ev
ent
burs
t
de
tec
t
io
n f
ro
m so
cia
l m
edia
t
e
x
t
Dha
rm
endra
M
a
ng
a
l
1
,
H
ema
nt
M
a
k
wa
na
2
1
D
e
p
a
r
t
me
n
t
o
f
C
S
E
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
M
e
d
i
-
C
a
p
s
U
n
i
v
e
r
si
t
y
,
I
n
d
o
r
e
,
I
n
d
i
a
2
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
V
,
I
n
d
o
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
Ju
n
6
,
2
0
2
4
R
ev
is
ed
Oct
17
,
2
0
2
4
Acc
ep
ted
Oct
30
,
2
0
2
4
Th
e
lan
g
u
a
g
e
m
o
d
e
ls
p
lay
v
e
ry
imp
o
rtan
t
ro
le
in
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
(NL
P
)
tas
k
s.
To
u
n
d
e
rsta
n
d
n
a
tu
ra
l
lan
g
u
a
g
e
s,
th
e
lea
rn
in
g
m
o
d
e
ls
a
re
re
q
u
ired
t
o
b
e
train
e
d
o
n
larg
e
c
o
r
p
u
s.
Th
is
re
q
u
ires
a
lo
t
o
f
ti
m
e
a
n
d
c
o
m
p
u
ti
n
g
re
so
u
rc
e
s.
Th
e
d
e
tec
ti
o
n
o
f
in
f
o
rm
a
ti
o
n
li
k
e
e
v
e
n
ts,
a
n
d
lo
c
a
ti
o
n
s
fr
o
m
tex
t
is
a
n
imp
o
rta
n
t
NLP
tas
k
.
As
e
v
e
n
ts
d
e
tec
ti
o
n
is
to
b
e
d
o
n
e
i
n
re
a
l
-
ti
m
e
so
th
a
t
imm
e
d
iate
a
c
ti
o
n
s
c
a
n
b
e
tak
e
n
,
h
e
n
c
e
we
n
e
e
d
e
fficie
n
t
d
e
c
isio
n
-
m
a
k
i
n
g
m
o
d
e
l
s.
Th
e
p
e
rtain
e
d
m
o
d
e
ls
li
k
e
b
i
-
d
irec
ti
o
n
a
l
e
n
c
o
d
e
rs
re
p
re
se
n
tatio
n
fr
o
m
tr
a
n
sfo
rm
e
rs
(BERT
)
g
a
in
in
g
p
o
p
u
larit
y
to
so
lv
e
NLP
p
r
o
b
lem
s.
As
BERT
b
a
se
d
m
o
d
e
ls
a
re
p
re
-
train
e
d
o
n
larg
e
lan
g
u
a
g
e
c
o
rp
u
s
it
re
q
u
ires
v
e
ry
l
e
ss
ti
m
e
to
a
d
a
p
t
fo
r
d
o
m
a
in
sp
e
c
ifi
c
NLP
tas
k
.
Diffe
re
n
t
imp
lem
e
n
tatio
n
s
o
f
BERT
h
a
v
e
b
e
e
n
p
r
o
p
o
se
d
t
o
e
n
h
a
n
c
e
e
fficie
n
c
y
a
n
d
a
p
p
li
c
a
b
il
it
y
o
f
th
e
b
a
se
m
o
d
e
l.
Th
e
se
lec
ti
o
n
o
f
rig
h
t
imp
lem
e
n
tatio
n
is
e
ss
e
n
ti
a
l
f
o
r
o
v
e
ra
ll
p
e
rfo
rm
a
n
c
e
o
f
NLP
b
a
se
d
sy
ste
m
.
Th
is
wo
r
k
p
re
se
n
ts
t
h
e
c
o
m
p
a
r
a
ti
v
e
in
si
g
h
ts
o
f
fi
v
e
wid
e
l
y
u
se
d
BER
T
imp
lem
e
n
tatio
n
s
n
a
m
e
d
a
s
BER
T
-
b
a
se
,
BERT
-
larg
e
,
Distil
l
BER
T,
Ro
b
u
st
BERT
a
p
p
r
o
a
c
h
(R
o
BERT
a
-
b
a
s
e
)
a
n
d
R
o
BERT
-
larg
e
fo
r
e
v
e
n
t
d
e
tec
ti
o
n
fro
m
th
e
tex
t
e
x
trac
ted
fro
m
s
o
c
ial
m
e
d
ia
stre
a
m
s.
Th
e
re
su
lt
s
sh
o
w
t
h
a
t
Distil
l
-
BERT
m
o
d
e
l
o
u
tp
e
rfo
r
m
s
o
n
b
a
sis
o
f
p
e
rfo
rm
a
n
c
e
m
e
tri
c
li
k
e
p
re
c
isio
n
,
re
c
a
ll
,
a
n
d
F
1
-
sc
o
re
w
h
il
e
th
e
fa
ste
st t
o
train
a
lso
.
K
ey
w
o
r
d
s
:
B
E
R
T
E
v
en
t d
etec
tio
n
Natu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
Per
f
o
r
m
an
ce
a
n
aly
s
is
So
cial
m
ed
ia
T
r
an
s
f
o
r
m
e
r
s
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
:
Dh
ar
m
en
d
r
a
Ma
n
g
al
Dep
ar
tm
en
t o
f
C
SE,
Facu
lty
o
f
E
n
g
in
ee
r
in
g
,
Me
d
i
-
C
ap
s
Un
iv
er
s
ity
I
n
d
o
r
e
,
I
n
d
ia
E
m
ail:
m
an
ag
ld
h
a
r
m
en
d
r
a8
3
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
s
o
cial
m
ed
ia
b
ec
o
m
es
th
e
in
f
o
r
m
atio
n
s
h
ar
in
g
p
latf
o
r
m
.
T
h
e
d
etec
tio
n
o
f
ev
en
ts
is
v
e
r
y
cr
u
cial
task
esp
ec
ially
f
o
r
th
at
n
ee
d
s
atten
tio
n
a
n
d
r
esp
o
n
s
e
b
y
ad
m
in
is
tr
ativ
e
o
f
f
icer
s
.
T
h
e
ev
en
ts
lik
e
f
lo
o
d
,
ea
r
th
q
u
ak
es,
r
io
ts
,
an
d
ac
cid
e
n
ts
n
ee
d
im
m
ed
iate
r
es
p
o
n
s
e
[
1
]
.
T
h
e
tim
ely
r
esp
o
n
s
e
r
eq
u
ir
es
tim
ely
d
etec
tio
n
.
T
h
is
h
elp
s
th
e
g
o
v
er
n
m
en
t
o
f
f
icials
to
r
ea
ct
o
n
co
r
r
e
ct
tim
e.
T
h
e
m
ess
ag
es
s
h
ar
ed
o
n
s
o
cial
m
ed
ia
ca
n
b
e
s
en
s
ed
to
d
etec
t
th
e
ev
en
ts
[
2
]
.
T
h
e
tex
t
m
ess
ag
es
ar
e
u
s
ed
as
in
p
u
t
to
d
etec
t
th
e
ev
en
ts
h
en
ce
ev
en
t
d
etec
tio
n
is
b
asically
a
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P
)
task
.
T
r
ad
itio
n
ally
NL
P
-
to
o
lk
it
(
NL
T
K
)
h
as
b
ee
n
u
s
ed
f
o
r
ev
en
t
d
etec
tio
n
as
it
is
v
er
y
s
im
p
le
to
u
s
e.
B
u
t
NL
T
K
r
elie
s
o
n
m
an
u
al
f
ea
tu
r
es
ex
tr
ac
tio
n
as
co
m
p
ar
ed
to
d
ee
p
lear
n
in
g
a
p
p
r
o
ac
h
es
wh
i
ch
p
r
o
v
id
es
au
t
o
m
atic
f
ea
tu
r
e
ex
tr
ac
tio
n
m
ec
h
an
is
m
[
3
]
.
C
o
n
s
eq
u
en
tly
,
r
ec
en
t
wo
r
k
s
ar
e
b
ased
o
n
d
ee
p
lear
n
in
g
m
o
d
els.
Alth
o
u
g
h
NL
T
K
is
u
s
ed
f
o
r
d
ata
p
r
ep
r
o
ce
s
s
in
g
.
Mo
s
t
o
f
th
e
d
ee
p
lear
n
in
g
m
o
d
els
-
b
ased
ap
p
r
o
ac
h
es
ar
e
r
ely
o
n
tex
t
r
e
p
r
es
en
tatio
n
f
ea
tu
r
e
s
lik
e
wo
r
d
e
m
b
ed
d
in
g
s
,
p
ar
t
-
of
-
s
p
ee
ch
t
ag
s
,
n
am
e
d
en
titi
es,
a
n
d
wo
r
d
p
o
s
itio
n
.
T
h
e
wo
r
d
e
m
b
ed
d
in
g
is
th
e
m
o
s
t
p
r
o
m
in
en
t
o
n
e
as
it
m
ap
s
in
p
u
t
to
k
en
(
wo
r
d
)
in
to
n
u
m
er
ic
v
alu
e.
T
h
is
is
m
an
d
ato
r
y
f
o
r
in
p
u
t
p
r
o
ce
s
s
in
g
b
y
a
n
y
d
ee
p
lear
n
in
g
m
o
d
el
as
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.
3
8
,
No
.
1
,
Ap
r
il
20
2
5
:
4
3
9
-
4
4
6
440
th
ey
ar
e
f
u
n
d
am
en
tally
m
at
h
em
atica
l
in
n
atu
r
e
[
4
]
.
Mo
s
t
o
f
th
e
d
ee
p
lear
n
in
g
m
et
h
o
d
s
p
r
o
p
o
s
ed
ar
e
co
n
v
o
l
u
tio
n
n
e
u
r
al
n
etwo
r
k
(
C
NN
)
an
d
r
ec
u
r
r
e
n
t n
eu
r
al
n
et
wo
r
k
(
R
NN
)
b
ased
[
5
]
.
T
h
e
f
ield
o
f
ev
en
t
d
etec
tio
n
attr
ac
ts
s
ig
n
if
ican
t
r
esear
ch
in
ter
est,
lead
in
g
to
th
e
p
u
b
lica
tio
n
o
f
n
u
m
er
o
u
s
r
esear
ch
p
ap
e
r
s
.
Var
io
u
s
r
esear
ch
g
r
o
u
p
s
em
p
l
o
y
d
if
f
er
e
n
t
an
o
m
aly
d
etec
tio
n
tech
n
iq
u
es,
NL
P
to
o
ls
,
m
o
d
alities
,
an
d
s
o
cial
n
etwo
r
k
s
,
f
o
c
u
s
in
g
o
n
d
iv
e
r
s
e
ap
p
licatio
n
s
[
6
]
.
Sev
e
r
al
s
u
r
v
ey
s
h
av
e
b
ee
n
co
n
d
u
cte
d
to
an
aly
ze
th
is
wea
lth
o
f
in
f
o
r
m
a
tio
n
.
Fo
r
e
x
am
p
l
e,
Z
h
o
u
et
a
l
.
[
7
]
co
n
d
u
cted
a
n
an
aly
s
is
o
f
s
o
cial
ev
en
t
d
etec
tio
n
ap
p
r
o
ac
h
es
f
r
o
m
a
m
o
d
ality
p
er
s
p
ec
tiv
e.
Ad
d
itio
n
ally
,
th
e
s
u
r
v
e
y
in
clu
d
es
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
s
o
f
v
ar
io
u
s
m
et
h
o
d
s
u
s
in
g
m
u
ltip
le
p
u
b
lic
d
a
tasets
.
W
h
ile
th
e
s
u
r
v
ey
d
o
e
s
n
o
t
p
r
io
r
itize
th
e
ev
alu
atio
n
p
r
o
ce
s
s
es
u
s
ed
in
r
ev
iewin
g
p
a
p
er
s
,
it
is
cr
u
cial
t
o
u
n
d
er
s
tan
d
h
o
w
au
th
o
r
s
g
at
h
er
an
d
lab
el
d
ata,
as we
ll a
s
th
e
m
etr
ics u
s
ed
to
ass
es
s
alg
o
r
ith
m
q
u
ality
f
r
o
m
a
co
m
p
ar
ativ
e
s
tan
d
p
o
in
t.
T
h
e
is
s
u
e
o
f
c
o
m
p
ar
i
n
g
e
v
en
t
d
etec
tio
n
ap
p
r
o
ac
h
es
is
em
p
h
asized
b
y
[
8
]
,
wh
o
also
p
r
esen
t
th
eir
p
r
o
p
o
s
al
f
o
r
ac
h
iev
in
g
r
ep
r
o
d
u
cib
le
r
esear
ch
in
ev
en
t
d
etec
tio
n
tech
n
iq
u
es
b
ased
o
n
T
witter
d
ata.
T
h
eir
m
ai
n
co
n
ce
p
t
in
v
o
l
v
es
cr
ea
tin
g
a
s
im
u
lated
T
witter
s
tr
ea
m
with
s
p
ec
if
ic
p
ar
am
eter
s
to
ass
ess
v
ar
io
u
s
m
eth
o
d
s
.
T
h
is
co
m
p
ar
is
o
n
en
ab
les
th
e
im
p
lem
en
tatio
n
o
f
d
if
f
er
e
n
t
ap
p
r
o
ac
h
es
with
in
th
e
s
am
e
en
v
ir
o
n
m
en
t.
Nev
er
th
eless
,
th
is
ap
p
r
o
ac
h
m
ay
n
o
t
b
e
ap
p
r
o
p
r
iate
f
o
r
s
o
lu
tio
n
s
th
at
r
ely
o
n
d
iv
e
r
s
e
d
ata
m
o
d
alities
d
u
e
to
th
e
ab
s
en
ce
o
f
m
etad
ata
f
o
r
ev
alu
atio
n
o
f
al
g
o
r
ith
m
q
u
ality
.
W
eiler
et
a
l.
[
9
]
,
[
1
0
]
h
a
v
e
a
u
th
o
r
ed
m
u
ltip
le
p
a
p
er
s
f
o
cu
s
in
g
o
n
an
i
n
-
d
ep
t
h
an
aly
s
is
o
f
m
etr
ics
s
u
itab
le
f
o
r
ev
alu
atio
n
.
T
h
e
y
r
ec
o
m
m
en
d
e
d
th
e
u
tili
za
tio
n
o
f
th
e
d
u
p
licate
ev
en
ts
r
ate
(
DE
R
ate)
to
ac
h
iev
e
a
m
o
r
e
p
r
ec
is
e
ev
alu
atio
n
o
f
m
eth
o
d
s
.
Fu
r
th
er
m
o
r
e,
i
n
c
o
n
s
id
er
atio
n
o
f
c
h
allen
g
es
as
s
o
ciate
d
with
d
ata
m
ar
k
u
p
,
th
e
au
th
o
r
s
s
u
g
g
ested
th
e
u
s
e
o
f
m
etr
ics
th
at
ca
n
b
e
a
u
to
m
atica
lly
o
b
tai
n
ed
.
Fo
r
in
s
tan
ce
,
q
u
an
titativ
e
m
etr
ics
lik
e
m
em
o
r
y
u
s
ag
e
o
r
ex
ec
u
tio
n
tim
e.
T
h
ey
also
p
r
o
p
o
s
ed
th
e
u
s
e
o
f
a
p
r
ec
is
io
n
m
etr
ic
b
ased
o
n
s
ea
r
ch
en
g
in
e
r
esu
lts
f
o
r
q
u
er
ies
r
elate
d
to
th
e
id
en
tifie
d
ev
e
n
ts
as
a
q
u
alita
tiv
e
m
ea
s
u
r
e.
T
h
e
in
co
r
p
o
r
atio
n
o
f
s
u
ch
a
m
etr
ic
en
ab
les a
m
o
r
e
eq
u
itab
le
c
o
m
p
ar
is
o
n
o
f
d
if
f
er
en
t m
eth
o
d
s
.
Ap
p
ly
in
g
tr
an
s
f
o
r
m
er
lik
e
B
E
R
T
f
o
r
ev
e
n
t
d
etec
tio
n
is
v
er
y
n
ew
id
ea
.
Ver
y
f
ew
r
esear
ch
er
s
h
av
e
b
ee
n
wo
r
k
ed
o
n
it.
T
h
er
e
ar
e
v
ar
io
u
s
B
E
R
T
im
p
lem
en
tati
o
n
s
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
.
T
h
e
Fig
u
r
e
1
s
h
o
ws
th
e
b
asic
ar
ch
itectu
r
e
o
f
B
E
R
T
b
ased
m
o
d
els.
T
h
e
B
E
R
T
b
ase
m
o
d
el
is
th
e
o
r
ig
in
o
f
all
t
h
e
d
if
f
e
r
en
t
B
E
R
T
im
p
lem
en
tatio
n
s
[
1
1
]
.
T
h
er
e
a
r
e
m
u
ltip
le
en
co
d
er
s
(
let
N)
p
r
esen
t in
B
E
R
T
b
ased
m
o
d
el
[
1
2
]
.
Fig
u
r
e
1
.
B
E
R
T
b
ased
m
o
d
el
g
en
er
al
ar
c
h
itectu
r
e
I
n
th
e
y
ea
r
2
0
1
7
n
ew
d
ee
p
le
ar
n
in
g
n
etwo
r
k
ar
ch
itectu
r
e
w
as
p
r
o
p
o
s
ed
b
y
Go
o
g
le
r
esear
ch
n
am
ed
as
tr
an
s
f
o
r
m
er
s
.
T
h
is
n
etwo
r
k
ar
ch
itectu
r
e
is
b
ased
o
n
th
e
i
d
ea
o
f
‘
atten
tio
n
’
b
ased
lear
n
in
g
.
T
h
e
‘
atten
tio
n
’
in
lear
n
in
g
m
ea
n
s
u
n
d
er
s
tan
d
in
g
t
h
e
wo
r
d
co
n
tex
t
in
b
o
t
h
th
e
d
ir
ec
tio
n
s
in
a
tex
t.
T
h
e
B
E
R
,
GPT
,
an
d
em
b
ed
d
in
g
s
f
r
o
m
lan
g
u
a
g
e
m
o
d
els
(
ELMo
)
a
r
e
p
r
im
a
r
ily
tr
an
s
f
o
r
m
er
m
o
d
els
[
1
3
]
,
[
1
4
]
.
W
o
r
d
em
b
ed
d
in
g
is
th
e
ce
n
t
r
al
f
ea
tu
r
e
in
m
o
s
t
o
f
th
e
N
L
P
task
[
1
5
]
.
T
h
e
atten
tio
n
m
ec
h
an
is
m
en
h
an
ce
s
th
e
w
o
r
d
e
m
b
ed
d
in
g
p
er
f
o
r
m
an
ce
.
Mo
s
t
o
f
th
e
p
r
e
v
io
u
s
wo
r
k
s
h
av
e
u
s
ed
wo
r
d
em
b
ed
d
in
g
m
o
d
els
lik
e
W
o
r
d
2
Vec
a
n
d
Glo
Ve.
B
u
t
p
r
e
-
tr
ai
n
ed
lan
g
u
ag
e
m
o
d
el
lik
e
B
E
R
T
h
as
p
r
o
v
e
d
i
ts
ab
ilit
y
to
p
r
o
v
id
e
b
etter
r
esu
lts
in
m
o
s
t
o
f
th
e
NL
P
task
s
u
ch
as
ev
e
n
t
d
et
ec
tio
n
f
o
r
m
te
x
t
m
ess
ag
e
[
1
6
]
.
T
h
e
B
E
R
T
is
a
tr
an
s
f
o
r
m
er
-
b
ased
m
o
d
el
wh
ich
f
o
llo
ws s
elf
-
s
u
p
er
v
is
ed
an
d
tr
an
s
f
er
lear
n
in
g
[
1
7
]
,
[
1
8
]
.
B
E
R
T
was c
r
ea
ted
to
jo
in
tly
tr
ain
th
e
lef
t
an
d
r
ig
h
t
co
n
tex
ts
in
o
r
d
er
to
p
r
e
-
t
r
ain
d
ee
p
b
id
ir
ec
tio
n
al
r
e
p
r
e
s
en
tatio
n
s
f
r
o
m
th
e
u
n
lab
eled
te
x
t.
I
ts
tr
ain
e
d
m
o
d
el
s
er
v
es
as
th
e
m
in
d
,
wh
ich
c
an
s
u
b
s
eq
u
e
n
tly
co
n
tr
o
l
a
r
e
th
e
in
cr
ea
s
in
g
ly
v
ast
co
llectio
n
s
d
is
co
v
er
ab
le
in
f
o
r
m
atio
n
an
d
q
u
e
r
ies
th
at
ar
e
t
ailo
r
ed
to
th
e
in
d
iv
id
u
al'
s
n
ee
d
s
.
T
h
is
p
r
o
ce
d
u
r
e
r
ef
er
r
ed
as
tr
an
s
f
er
lear
n
in
g
[
1
9
]
.
T
h
e
B
E
R
T
m
o
d
el
p
r
e
-
tr
ain
ed
o
n
a
v
ast
co
llectio
n
o
f
u
n
la
b
eled
tex
t
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
P
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f d
iffer
en
t B
E
R
T imp
leme
n
ta
tio
n
fo
r
ev
en
t
…
(
Dh
a
r
men
d
r
a
Ma
n
g
a
l
)
441
en
co
m
p
ass
in
g
W
ik
ip
ed
ia
(
2
,
5
0
0
m
illi
o
n
wo
r
d
s
)
an
d
liter
a
r
y
wo
r
k
s
(
E
n
g
lis
h
)
.
Du
e
to
th
e
e
x
ten
s
iv
e
tr
ain
in
g
o
n
a
lar
g
e
tex
t
co
r
p
u
s
,
th
e
m
o
d
e
l
g
r
ad
u
ally
d
ev
elo
p
s
a
co
m
p
r
eh
en
s
iv
e
an
d
th
o
r
o
u
g
h
u
n
d
er
s
tan
d
in
g
o
f
E
n
g
lis
h
lan
g
u
ag
e
[
2
0
]
.
2.
M
E
T
H
O
D
T
h
e
n
o
tio
n
o
f
tr
an
s
f
o
r
m
er
in
NL
P
is
r
ev
o
lu
tio
n
ar
y
o
n
e
.
T
h
e
tr
an
s
f
o
r
m
er
s
ar
e
b
asically
d
e
ep
lear
n
i
n
g
m
o
d
el
-
b
ased
s
y
s
tem
s
with
att
en
tio
n
[
2
1
]
.
All
B
E
R
T
im
p
lem
en
tatio
n
s
i.e
.
B
E
R
T
-
b
ase,
B
E
R
T
-
lar
g
e,
Dis
till
-
B
E
R
T
,
R
o
B
E
R
T
a
-
b
ase
,
an
d
R
o
B
E
R
T
a
-
lar
g
e
ar
e
co
m
p
ar
e
d
in
th
is
wo
r
k
.
T
h
e
b
r
ief
d
e
s
cr
ip
tio
n
o
f
t
h
ese
m
o
d
els
is
g
iv
en
in
f
o
llo
win
g
p
a
r
ag
r
ap
h
s
.
T
h
e
B
E
R
T
m
o
d
el
p
r
o
ce
s
s
5
1
2
to
k
en
s
in
p
u
t
an
d
o
u
tp
u
ts
t
h
e
v
ec
to
r
r
ep
r
esen
tatio
n
o
f
th
e
s
eq
u
en
ce
i.e
.
wo
r
d
e
m
b
ed
d
in
g
[
2
2
]
.
T
h
is
o
u
tp
u
t
ca
n
h
av
e
o
n
e
o
r
two
s
eg
m
en
ts
,
with
t
h
e
f
ir
s
t
to
k
en
a
lway
s
b
ein
g
[
C
L
S],
co
n
tain
in
g
th
e
s
p
ec
if
ic
class
if
icatio
n
em
b
e
d
d
in
g
,
an
d
an
o
th
er
s
p
ec
ial
t
o
k
en
,
[
SEP]
,
u
s
ed
to
s
ep
ar
ate
th
e
s
eg
m
en
ts
.
B
E
R
T
o
r
g
an
izes
th
e
f
in
al
h
i
d
d
en
s
tate
h
o
f
th
e
f
i
r
s
t
to
k
en
[
C
L
S]
in
o
r
d
er
t
o
p
r
o
ce
s
s
th
e
co
m
p
lete
s
eq
u
en
ce
f
o
r
te
x
t
class
if
icatio
n
task
s
.
T
o
o
b
tain
th
e
p
r
e
d
icte
d
p
r
o
b
ab
ilit
ies
f
r
o
m
th
e
tr
ain
e
d
m
o
d
el
a
So
f
tMa
x
class
if
ier
is
in
clu
d
ed
at
th
e
to
p
o
f
th
e
B
E
R
T
m
o
d
el
[
2
3
]
.
T
h
e
d
ata
s
et
n
ee
d
s
to
b
e
co
n
v
er
ted
in
t
o
v
ec
to
r
s
b
ef
o
r
e
b
ei
n
g
in
p
u
t
in
to
th
e
cla
s
s
if
ier
b
ec
au
s
e
it
is
in
itially
in
tex
t
f
o
r
m
.
B
E
R
T
lear
n
s
co
n
tex
tu
al
em
b
ed
d
i
n
g
s
in
s
tead
o
f
c
o
n
tex
t
-
f
r
ee
em
b
e
d
d
in
g
s
,
u
n
lik
e
W
o
r
d
2
Vec
.
E
v
en
th
o
u
g
h
th
er
e
ar
e
d
if
f
er
e
n
t
m
o
d
els
f
o
r
te
x
t
v
ec
to
r
izatio
n
,
B
E
R
T
ca
r
r
ies
o
u
t
to
k
en
izatio
n
u
s
in
g
th
e
W
o
r
d
Piece
ap
p
r
o
ac
h
.
T
h
e
f
o
u
n
d
atio
n
o
f
B
E
R
T
is
a
s
tack
o
f
en
co
d
er
la
y
er
s
.
T
h
e
n
u
m
b
er
o
f
en
c
o
d
er
lay
er
s
is
w
h
er
e
B
E
R
T
b
ase
a
n
d
B
E
R
T
-
lar
g
e
d
iv
er
g
e.
I
n
th
e
B
E
R
T
-
lar
g
e
m
o
d
el,
th
er
e
ar
e
2
4
lay
er
s
o
f
en
co
d
e
r
s
lay
er
ed
o
n
to
p
o
f
o
n
e
an
o
th
er
,
co
m
p
ar
ed
to
1
2
lay
er
s
in
th
e
B
E
R
T
b
ase
m
o
d
el
[
2
4
]
.
T
h
e
Dis
til
B
E
R
T
d
ec
r
ea
s
es
th
e
s
ize
o
f
B
E
R
T
b
y
4
0
%
wh
ile
k
ee
p
in
g
9
7
%
o
f
B
E
R
T
's
p
er
f
o
r
m
a
n
ce
[
2
5
]
.
Dis
tilB
E
R
T
r
em
o
v
es
p
o
o
ler
a
n
d
to
k
en
-
ty
p
e
em
b
ed
d
i
n
g
s
to
r
esem
b
le
th
e
B
E
R
T
m
o
d
el.
Dis
till
a
tio
n
is
th
e
tech
n
iq
u
e
o
f
ap
p
r
o
x
im
atin
g
a
lar
g
er
n
e
two
r
k
'
s
f
u
ll
o
u
tp
u
t
d
is
tr
ib
u
tio
n
s
u
s
in
g
a
s
m
aller
n
etwo
r
k
af
ter
th
e
lar
g
er
n
etw
o
r
k
h
as
b
ee
n
tr
ai
n
ed
.
T
h
is
is
an
alo
g
o
u
s
to
p
o
s
ter
io
r
ap
p
r
o
x
im
atio
n
in
ce
r
tain
r
esp
ec
ts
,
an
d
Ku
lb
ac
k
L
eib
er
d
iv
er
g
en
ce
is
o
n
e
o
f
th
e
k
e
y
o
p
tim
izatio
n
p
r
o
c
ed
u
r
es a
p
p
lie
d
[
2
6
]
.
R
o
B
E
R
T
a,
s
h
o
r
t
f
o
r
r
o
b
u
s
tly
o
p
tim
ized
B
E
R
T
ap
p
r
o
ac
h
,
was
in
tr
o
d
u
ce
d
b
y
Face
b
o
o
k
[
2
7
]
.
I
t
in
v
o
lv
es r
etr
ain
in
g
B
E
R
T
u
s
in
g
an
en
h
a
n
ce
d
tr
ain
in
g
m
eth
o
d
o
lo
g
y
,
a
lar
g
er
d
ataset,
an
d
in
cr
ea
s
ed
co
m
p
u
tin
g
p
o
wer
.
T
h
e
R
o
B
E
R
T
a
m
o
d
e
l
is
im
p
lem
en
ted
s
im
ilar
to
th
e
B
E
R
T
m
o
d
el,
with
a
m
i
n
o
r
ch
an
g
e
to
th
e
em
b
ed
d
in
g
s
an
d
a
s
etu
p
f
o
r
p
r
etr
ain
in
g
R
o
B
E
R
T
a
m
o
d
els.
Alth
o
u
g
h
it sh
ar
es th
e
s
am
e
ar
ch
itectu
r
e
as B
E
R
T
,
R
o
B
E
R
T
a
u
tili
ze
s
a
b
y
te
-
lev
el
p
air
en
co
d
in
g
(
B
PE)
to
k
e
n
izer
s
im
ilar
to
GPT
-
2
a
n
d
em
p
lo
y
s
a
d
i
f
f
er
en
t
p
r
etr
ain
in
g
s
ch
em
e.
L
ar
g
e
m
in
i
-
b
atch
es,
a
h
ig
h
er
b
y
te
-
le
v
el
B
PE,
d
y
n
am
ic
m
ask
i
n
g
,
an
d
e
n
tire
p
h
r
ases
with
o
u
t
NSP
lo
s
s
ar
e
all
u
s
ed
in
th
e
tr
ain
in
g
o
f
R
o
B
E
R
T
a.
B
y
elim
in
atin
g
th
e
n
ex
t
s
en
ten
ce
p
r
ed
ictio
n
(
NSP)
task
f
r
o
m
B
E
R
T
's
p
r
e
-
tr
ain
in
g
an
d
im
p
lem
e
n
tin
g
d
y
n
a
m
ic
m
ask
in
g
to
alter
t
h
e
m
ask
e
d
t
o
k
en
d
u
r
i
n
g
tr
ain
in
g
ep
o
ch
s
,
R
o
B
E
R
T
a
en
h
an
ce
s
th
e
tr
ain
in
g
p
r
o
ce
s
s
[
2
8
]
.
T
h
e
e
x
p
er
im
en
t d
e
m
o
n
s
tr
ated
th
at
l
ar
g
er
b
atch
tr
ain
i
n
g
s
izes
ar
e
m
o
r
e
b
en
ef
icial
i
n
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
.
No
tab
ly
,
in
ad
d
itio
n
t
o
B
E
R
T
tr
ain
in
g
1
6
GB
o
f
b
o
o
k
s
C
o
r
p
u
s
an
d
E
n
g
lis
h
W
ik
ip
e
d
ia
d
ata,
R
o
B
E
R
T
a
u
tili
ze
s
1
6
0
GB
o
f
te
x
t
f
o
r
p
r
e
-
tr
ain
in
g
.
I
n
t
h
e
R
o
B
E
R
T
a
-
lar
g
e
m
o
d
el,
th
er
e
ar
e
2
4
lay
er
s
o
f
en
co
d
er
s
lay
er
ed
o
n
to
p
o
f
o
n
e
an
o
th
er
,
co
m
p
ar
ed
t
o
1
2
lay
er
s
in
th
e
R
o
B
E
R
T
a
b
ase
m
o
d
el
[
2
9
]
.
T
h
e
F
ig
u
r
e
2
s
h
o
ws th
e
m
eth
o
d
o
lo
g
y
a
d
ap
ted
in
th
is
wo
r
k
.
Af
ter
p
r
ep
r
o
ce
s
s
in
g
o
f
th
e
d
ataset
,
all
th
e
f
iv
e
m
o
d
els
ar
e
ap
p
lied
s
eq
u
en
tially
.
Fin
ally
,
th
ei
r
p
e
r
f
o
r
m
an
ce
is
co
m
p
ar
ed
.
T
h
e
d
e
tailed
ar
ch
itectu
r
al
d
if
f
er
en
ce
s
a
m
o
n
g
all
f
i
v
e
m
o
d
els
ar
e
s
h
o
wn
in
T
ab
le
1
.
I
t
s
h
o
ws
co
m
p
ar
is
o
n
am
o
n
g
t
h
e
co
n
s
id
er
ed
B
E
R
T
v
ar
ian
ts
b
ased
o
n
ch
ar
ac
te
r
is
tics
lik
e
n
u
m
b
er
o
f
en
co
d
er
s
,
h
id
d
en
lay
er
s
,
s
elf
-
atten
tio
n
h
ea
d
s
an
d
d
ec
is
io
n
p
ar
am
eter
s
.
Fig
u
r
e
2
.
Me
th
o
d
o
lo
g
y
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.
3
8
,
No
.
1
,
Ap
r
il
20
2
5
:
4
3
9
-
4
4
6
442
T
ab
le
1
.
Ar
c
h
itectu
r
al
d
if
f
e
r
en
ce
s
am
o
n
g
f
i
v
e
d
if
f
e
r
en
t BER
T
-
im
p
lem
en
tatio
n
s
M
o
d
e
l
#
E
n
c
o
d
e
r
s (N
)
H
i
d
d
e
n
l
a
y
e
r
s
i
z
e
#
S
e
l
f
a
t
t
e
n
t
i
o
n
h
e
a
d
s
#
P
a
r
a
me
t
e
r
s
B
ER
T
-
b
a
se
12
7
6
8
12
1
1
0
M
B
ER
T
-
l
a
r
g
e
24
1
0
2
4
16
3
4
0
M
D
i
st
i
l
l
-
B
ER
T
6
7
6
8
22
6
6
M
R
o
B
E
R
Ta
-
b
a
s
e
12
7
6
8
12
1
2
5
M
R
O
B
E
R
Ta
-
l
a
r
g
e
24
1
0
2
4
16
3
5
5
M
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
ex
p
e
r
im
en
t
d
ataset
co
n
tain
s
m
o
r
e
th
an
1
0
K
te
x
t
m
ess
ag
es
tak
en
f
r
o
m
X
(
T
witter
)
p
latf
o
r
m
.
T
h
is
d
ata
s
et
i
s
d
iv
id
ed
in
to
two
p
ar
ts
:
tr
ain
in
g
s
et
(
8
5
%
)
an
d
test
in
g
s
et
(
1
5
%).
T
h
e
tr
ain
in
g
o
f
all
th
e
ca
n
d
id
ate
m
o
d
els
h
as
b
ee
n
p
er
f
o
r
m
e
d
in
1
0
ep
o
ch
s
.
Als
o
,
ADAM
o
p
tim
izer
is
u
s
ed
f
o
r
h
y
p
er
-
p
ar
am
ete
r
tu
n
in
g
.
T
h
e
h
y
p
er
-
p
ar
am
ete
r
s
in
clu
d
e
ac
c
u
r
ac
y
an
d
lo
s
s
v
alu
e.
Factu
ally
,
ac
cu
r
ac
y
is
a
m
etr
ic
th
at
d
escr
ib
e
p
er
ce
n
tag
e
o
f
th
e
test
o
r
v
alid
atio
n
d
ata
co
r
r
ec
tly
lab
ele
d
wh
er
ea
s
lo
s
s
v
alu
e
is
th
e
av
er
ag
e
d
is
tan
ce
b
etwe
en
th
e
tr
u
e
v
al
u
es
an
d
th
e
v
alu
e
s
p
r
ed
icted
b
y
th
e
m
o
d
el.
T
h
e
d
atasets
u
s
ed
co
m
p
r
is
es
m
o
r
e
th
an
1
0
K
tex
t
me
s
s
ag
es
(
twee
ts
)
av
ailab
le
f
r
o
m
Kag
g
le
p
latf
o
r
m
.
T
h
e
F
ig
u
r
e
3
to
Fig
u
r
e
7
s
h
o
ws
th
e
lear
n
in
g
cu
r
v
es
f
o
r
d
if
f
er
en
t BER
T
m
o
d
els f
o
r
ac
cu
r
ac
y
an
d
lo
s
s
v
alu
e
r
esp
ec
ti
v
ely
.
Fig
u
r
e
3
.
Hy
p
er
-
p
ar
am
eter
s
(
a
cc
u
r
ac
y
an
d
lo
s
s
)
tu
n
in
g
f
o
r
B
E
R
T
-
b
ase
m
o
d
el
in
1
0
ep
o
ch
s
Fig
u
r
e
4
.
Hy
p
er
-
p
ar
am
eter
s
(
a
cc
u
r
ac
y
an
d
lo
s
s
)
tu
n
in
g
f
o
r
B
E
R
T
-
lar
g
e
m
o
d
el
in
1
0
e
p
o
ch
s
Fig
u
r
e
5
.
Hy
p
er
-
p
ar
am
eter
s
(
a
cc
u
r
ac
y
an
d
lo
s
s
)
tu
n
in
g
f
o
r
d
i
s
till
-
B
E
R
T
m
o
d
el
in
1
0
e
p
o
ch
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
P
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f d
iffer
en
t B
E
R
T imp
leme
n
ta
tio
n
fo
r
ev
en
t
…
(
Dh
a
r
men
d
r
a
Ma
n
g
a
l
)
443
Fig
u
r
e
6
.
Hy
p
er
-
p
ar
am
eter
s
(
a
cc
u
r
ac
y
an
d
lo
s
s
)
tu
n
in
g
f
o
r
R
o
B
E
R
T
a
-
b
ase
m
o
d
el
in
1
0
ep
o
ch
s
Fig
u
r
e
7
.
Hy
p
er
-
p
ar
am
eter
s
(
a
cc
u
r
ac
y
an
d
lo
s
s
)
tu
n
in
g
f
o
r
R
o
B
E
R
T
a
-
b
ase
m
o
d
el
in
1
0
ep
o
ch
s
I
n
s
u
m
m
ar
y
,
th
e
Fig
u
r
e
8
an
d
9
s
h
o
ws
th
e
p
er
f
o
r
m
an
ce
c
o
m
p
ar
is
o
n
am
o
n
g
all
th
e
f
iv
e
m
o
d
el
d
u
r
in
g
f
in
e
tu
n
in
g
in
1
0
iter
atio
n
(
ep
o
ch
s
)
o
f
lear
n
in
g
.
T
h
er
e
a
r
e
3
p
er
f
o
r
m
an
ce
m
etr
ics
ar
e
u
s
ed
f
o
r
c
o
m
p
a
r
is
o
n
i.e
.
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F
1
-
s
co
r
e
.
Fig
u
r
e
1
0
s
h
o
ws th
e
r
elatio
n
am
o
n
g
t
h
em
.
Fig
u
r
e
8
.
C
o
m
p
a
r
is
o
n
o
f
ac
cu
r
ac
y
d
u
r
i
n
g
f
i
n
e
tu
n
in
g
in
ea
c
h
ep
o
ch
Fig
u
r
e
9
.
C
o
m
p
a
r
is
o
n
o
f
lo
s
s
-
v
alu
e
d
u
r
in
g
f
in
e
tu
n
i
n
g
in
ea
c
h
ep
o
c
h
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.
3
8
,
No
.
1
,
Ap
r
il
20
2
5
:
4
3
9
-
4
4
6
444
Fig
u
r
e
1
0
.
Der
iv
in
g
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
-
s
co
r
e
f
r
o
m
r
aw
m
ea
s
u
r
em
en
ts
As
we
ca
n
co
n
s
id
er
th
e
ev
en
t
b
u
r
s
t
id
en
tific
atio
n
as
b
in
ar
y
class
if
icat
io
n
p
r
o
b
lem
i.e
.
w
h
eth
er
th
e
tex
t
m
ess
ag
e
in
d
icatin
g
an
ev
en
t
b
u
r
s
t
(
p
o
s
itiv
e
+)
o
r
n
o
t
(
n
eg
ativ
e
-
)
.
T
h
u
s
,
th
e
m
o
d
el
p
r
ed
ictio
n
s
ar
e
b
ased
o
n
th
ese
b
in
ar
y
class
es.
T
h
e
tr
ain
in
g
d
ata
is
lab
eled
with
b
in
ar
y
lab
els
(
1
f
o
r
p
o
s
itiv
e
an
d
0
f
o
r
n
eg
ativ
e)
.
T
r
u
e
p
o
s
itiv
es
(
tr
u
e+
)
is
co
u
n
t
o
f
co
r
r
ec
tly
id
e
n
tifie
d
p
o
s
itiv
e
lab
els
wh
er
e
as
f
alse
p
o
s
itiv
e
(
f
alse+)
in
d
icate
s
co
u
n
t
o
f
p
o
s
itiv
e
lab
el
class
if
ied
in
co
r
r
ec
tly
.
Similar
ly
tr
u
e
n
eg
ativ
es
(
tr
u
e
-
)
ar
e
th
e
co
u
n
t
o
f
co
r
r
ec
tly
id
en
tifie
d
n
eg
ativ
e
lab
els
wh
e
r
ea
s
f
alse
n
eg
ativ
es(f
alse
-
)
ar
e
co
u
n
t
o
f
in
co
r
r
ec
tly
class
if
ied
n
eg
ativ
e
lab
els
b
y
th
e
m
o
d
el.
W
e
ca
n
d
er
iv
e
t
h
e
v
alu
es o
f
p
r
ec
is
io
n
an
d
r
ec
all
u
s
in
g
th
e
f
o
ll
o
win
g
f
o
r
m
u
las:
=
(
+
)
/
(
+
)
+
(
+
)
=
(
+
)
/
(
+
)
+
(
−
)
T
h
e
F1
-
s
co
r
e
is
th
e
h
a
r
m
o
n
ic
m
ea
n
o
f
p
r
ec
is
io
n
an
d
r
ec
all
v
alu
e.
I
t is
f
o
r
m
u
lated
as
:
1
−
=
(
2
∗
∗
)
/
+
)
T
h
ese
m
etr
ics
h
as
b
ee
n
co
m
p
u
ted
f
o
r
d
i
f
f
er
en
t
B
E
R
T
v
ar
ian
ts
as
s
u
m
m
ar
ized
in
T
ab
le
2
.
I
t
s
h
o
ws
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F1
v
alu
e
o
f
th
e
co
m
p
ar
e
d
B
E
R
T
v
ar
ian
ts
f
o
r
b
o
th
p
o
s
itiv
e
lab
e
led
a
n
d
n
eg
ativ
e
lab
e
led
d
ata
in
th
e
d
ataset.
T
h
is
s
tu
d
y
is
u
n
iq
u
e
a
n
d
co
m
p
r
eh
e
n
s
iv
e
o
n
e.
T
a
b
le
3
s
h
o
ws
th
e
d
if
f
er
en
ce
s
with
s
o
m
e
p
r
ev
io
u
s
wo
r
k
.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
d
if
f
e
r
en
t BER
T
v
ar
ian
ts
b
ased
o
n
p
er
f
o
r
m
a
n
ce
m
etr
ics
M
o
d
e
l
P
r
e
c
i
s
i
o
n
(
l
a
b
e
l
0
)
P
r
e
c
i
s
i
o
n
(
l
a
b
e
l
1
)
R
e
c
a
l
l
(
l
a
b
e
l
0
)
R
e
c
a
l
l
(
l
a
b
e
l
0
)
F
1
-
sc
o
r
e
(
l
a
b
e
l
0)
F
1
-
sc
o
r
e
(
l
a
b
e
l
1
)
B
ER
T
-
b
a
se
0
.
8
8
0
.
9
3
0
.
9
6
0
.
8
2
0
.
9
1
0
.
8
7
B
ER
T
-
l
a
r
g
e
0
.
8
8
0
.
9
4
0
.
9
6
0
.
8
3
0
.
9
2
0
.
8
8
D
i
st
i
l
l
-
B
ER
T
0
.
9
0
.
9
5
0
.
9
6
0
.
8
7
0
.
9
3
0
.
9
1
R
o
B
E
R
Ta
-
b
a
s
e
0
.
8
9
0
.
9
5
0
.
9
7
0
.
8
5
0
.
9
3
0
.
9
R
O
B
E
R
Ta
-
l
a
r
g
e
0
.
8
9
0
.
8
5
0
.
8
8
0
.
8
5
0
.
8
9
0
.
8
5
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
with
p
r
e
v
io
u
s
wo
r
k
A
u
t
h
o
r
(
s)
M
o
d
e
l
s
c
o
mp
a
r
e
d
C
o
r
p
u
s
u
s
e
d
F
i
n
d
i
n
g
s
Tu
r
c
h
i
n
et
a
l
.
[
2
0
]
B
ER
T
b
a
se
,
c
l
i
n
i
c
a
l
B
E
R
T
,
a
n
d
B
i
o
B
ER
T
M
e
d
i
c
a
l
t
e
x
t
d
o
c
u
me
n
t
s
C
l
i
n
i
c
a
l
B
E
R
T
o
u
t
p
e
r
f
o
r
ms
.
Ze
i
n
a
l
i
et
al
.
[
2
3
]
B
ER
T
b
a
se
,
s
p
a
n
B
E
R
T,
c
l
i
n
i
c
a
l
B
E
R
T
,
a
n
d
B
i
o
B
E
R
T
El
e
c
t
r
o
n
i
c
h
e
a
l
t
h
r
e
c
o
r
d
s
C
l
i
n
i
c
a
l
B
E
R
T
o
u
t
p
e
r
f
o
r
ms
.
C
o
r
t
i
z
[
2
4
]
B
ER
T
b
a
se
,
D
i
st
i
l
-
B
ER
T,
R
o
B
E
R
T
a
,
X
LN
e
t
,
a
n
d
ELE
C
T
R
A
Emo
t
i
o
n
’
s
d
a
t
a
set
s
D
i
st
i
l
B
ER
T
o
u
t
p
e
r
f
o
r
ms
P
r
o
p
o
se
d
w
o
r
k
B
ER
T
b
a
se
/
l
a
r
g
e
,
D
i
st
i
l
-
B
ER
T
,
a
n
d
R
o
B
E
R
Ta
b
a
se
/
l
a
r
g
e
D
i
sast
e
r
e
v
e
n
t
t
w
e
e
t
s
D
i
st
i
l
B
ER
T
a
n
d
R
o
B
ER
Ta
b
a
s
e
o
u
t
p
e
r
f
o
r
ms.
4.
CO
NCLU
SI
O
N
Use
o
f
p
r
e
-
tr
ain
e
d
lan
g
u
ag
e
m
o
d
els
in
NL
P
ap
p
licatio
n
s
led
to
s
ig
n
if
ican
t
p
er
f
o
r
m
an
ce
g
ain
.
As
v
ar
io
u
s
p
r
e
-
tr
ain
e
d
m
o
d
els
a
r
e
av
ailab
le
ca
r
ef
u
l
co
m
p
ar
is
o
n
b
etwe
en
th
em
is
a
ch
allen
g
i
n
g
task
.
T
h
is
h
elp
s
th
e
NL
P
s
o
lu
tio
n
ar
c
h
itect
to
ch
o
o
s
e
th
e
m
o
s
t
ap
p
r
o
p
r
iate
a
m
o
n
g
t
h
e
all
av
ailab
le.
I
n
th
is
wo
r
k
we
c
ar
ef
u
lly
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
P
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f d
iffer
en
t B
E
R
T imp
leme
n
ta
tio
n
fo
r
ev
en
t
…
(
Dh
a
r
men
d
r
a
Ma
n
g
a
l
)
445
o
b
s
er
v
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
d
i
f
f
er
en
t
B
E
R
T
im
p
lem
en
tatio
n
s
o
n
ev
e
n
t
d
etec
tio
n
task
f
r
o
m
s
o
cial
m
ed
ia
tex
t.
T
h
e
f
iv
e
p
o
p
u
la
r
B
E
R
T
im
p
lem
en
tatio
n
s
n
am
ely
B
E
R
T
-
b
ase,
B
E
R
T
-
lar
g
e,
Dis
til
l
-
B
E
R
T
,
R
o
B
E
R
T
a
–
b
ase
,
an
d
R
o
B
E
R
T
a
-
lar
g
e
ar
e
co
m
p
ar
ed
b
ased
o
n
th
e
p
e
r
f
o
r
m
a
n
ce
m
etr
ics
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F
1
-
s
co
r
e.
I
n
co
n
clu
s
io
n
,
we
h
av
e
f
o
u
n
d
th
at
Dis
till
-
B
E
R
T
im
p
lem
en
tatio
n
tr
ain
e
d
o
n
e
v
en
t
d
etec
tio
n
d
ataset
o
u
tp
er
f
o
r
m
s
am
o
n
g
all
o
th
er
wh
ile
R
o
B
E
R
T
a
-
b
ase
m
o
d
el
p
er
f
o
r
m
ed
i
m
p
r
ess
iv
e
alm
o
s
t
eq
u
al
to
Dis
till
-
B
E
R
T
m
o
d
el.
T
h
is
s
tu
d
y
f
u
r
th
e
r
ex
ten
d
ed
f
o
r
ar
ea
s
lik
e
m
e
d
ical
d
o
c
u
m
en
t
an
aly
s
is
,
tr
av
el
b
lo
g
an
aly
s
is
f
o
r
d
is
ea
s
e
p
r
ed
ictio
n
a
n
d
s
p
atial
in
f
o
r
m
a
tio
n
ex
tr
ac
tio
n
r
esp
ec
tiv
ely
.
RE
F
E
R
E
NC
E
S
[
1
]
M
.
D
i
n
g
,
C
.
Z
h
o
u
,
H
.
Y
a
n
g
,
a
n
d
J.
T
a
n
g
,
“
C
o
g
LTX
:
a
p
p
l
y
i
n
g
B
ER
T
t
o
l
o
n
g
t
e
x
t
s
,
”
Ad
v
a
n
c
e
s
i
n
N
e
u
ra
l
I
n
f
o
rm
a
t
i
o
n
Pr
o
c
e
ssi
n
g
S
y
s
t
e
m
s
,
v
o
l
.
2
0
2
0
-
D
e
c
e
mb
e
r
,
2
0
2
0
.
[
2
]
M
.
Za
h
e
e
r
e
t
a
l
.
,
“
B
i
g
b
i
r
d
:
t
r
a
n
sf
o
r
me
r
s
f
o
r
l
o
n
g
e
r
se
q
u
e
n
c
e
s
,
”
A
d
v
a
n
c
e
s
i
n
N
e
u
r
a
l
I
n
f
o
rm
a
t
i
o
n
Pro
c
e
ss
i
n
g
S
y
st
e
m
s
,
v
o
l
.
2
0
2
0
-
D
e
c
e
m
b
e
r
,
2
0
2
0
.
[3
]
C
.
C
a
su
l
a
a
n
d
S
.
T
o
n
e
l
l
i
,
“
H
a
t
e
s
p
e
e
c
h
d
e
t
e
c
t
i
o
n
w
i
t
h
mac
h
i
n
e
-
t
r
a
n
sl
a
t
e
d
d
a
t
a
:
t
h
e
r
o
l
e
o
f
a
n
n
o
t
a
t
i
o
n
s
c
h
e
me
,
c
l
a
ss i
mb
a
l
a
n
c
e
a
n
d
u
n
d
e
r
s
a
mp
l
i
n
g
,
”
C
EU
R
Wo
r
k
sh
o
p
Pr
o
c
e
e
d
i
n
g
s
,
v
o
l
.
2
7
6
9
,
2
0
2
0
,
d
o
i
:
1
0
.
4
0
0
0
/
b
o
o
k
s.a
a
c
c
a
d
e
m
i
a
.
8
3
4
5
.
[
4
]
M
.
B
u
d
a
,
A
.
M
a
k
i
,
a
n
d
M
.
A
.
M
a
z
u
r
o
w
sk
i
,
“
A
sy
s
t
e
m
a
t
i
c
s
t
u
d
y
o
f
t
h
e
c
l
a
ss
i
m
b
a
l
a
n
c
e
p
r
o
b
l
e
m
i
n
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
N
e
u
r
a
l
N
e
t
w
o
r
k
s
,
v
o
l
.
1
0
6
,
p
p
.
2
4
9
–
2
5
9
,
2
0
1
8
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
n
e
t
.
2
0
1
8
.
0
7
.
0
1
1
.
[
5
]
A
.
V
a
sw
a
n
i
e
t
a
l
.
,
“
A
t
t
e
n
t
i
o
n
i
s
a
l
l
y
o
u
n
e
e
d
i
n
a
d
v
a
n
c
e
s
i
n
n
e
u
r
a
l
i
n
f
o
r
m
a
t
i
o
n
p
r
o
c
e
ssi
n
g
s
y
st
e
ms,
”
S
e
a
rc
h
Pu
b
Me
d
,
p
p
.
5
9
9
8
–
6
0
0
8
,
2
0
1
7
.
[
6
]
Y
.
Zh
u
e
t
a
l
.
,
“
A
l
i
g
n
i
n
g
b
o
o
k
s
a
n
d
mo
v
i
e
s
:
t
o
w
a
r
d
s
st
o
r
y
-
l
i
k
e
v
i
s
u
a
l
e
x
p
l
a
n
a
t
i
o
n
s
b
y
w
a
t
c
h
i
n
g
m
o
v
i
e
s
a
n
d
r
e
a
d
i
n
g
b
o
o
k
s,”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
,
v
o
l
.
2
0
1
5
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
mp
u
t
e
r
V
i
s
i
o
n
,
I
C
C
V
2
0
1
5
,
p
p
.
1
9
–
2
7
,
2
0
1
5
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
V
.
2
0
1
5
.
1
1
.
[
7
]
H
.
Z
h
o
u
,
H
.
Y
i
n
,
H
.
Zh
e
n
g
,
a
n
d
Y
.
L
i
,
“
A
su
r
v
e
y
o
n
mu
l
t
i
-
mo
d
a
l
so
c
i
a
l
e
v
e
n
t
d
e
t
e
c
t
i
o
n
,
”
K
n
o
w
l
e
d
g
e
-
B
a
se
d
S
y
st
e
m
s
,
v
o
l
.
1
9
5
,
p
.
1
0
5
6
9
5
,
M
a
y
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
k
n
o
sy
s
.
2
0
2
0
.
1
0
5
6
9
5
.
[
8
]
A
.
W
e
i
l
e
r
,
H
.
S
c
h
i
l
l
i
n
g
,
L
.
K
i
r
c
h
e
r
,
a
n
d
M
.
G
r
o
ssn
i
k
l
a
u
s
,
“
To
w
a
r
d
s
r
e
p
r
o
d
u
c
i
b
l
e
r
e
sea
r
c
h
o
f
e
v
e
n
t
d
e
t
e
c
t
i
o
n
t
e
c
h
n
i
q
u
e
s
f
o
r
Tw
i
t
t
e
r
,
”
i
n
Pro
c
e
e
d
i
n
g
s
-
6
t
h
S
w
i
ss
C
o
n
f
e
re
n
c
e
o
n
D
a
t
a
S
c
i
e
n
c
e
,
S
D
S
2
0
1
9
,
I
EEE,
J
u
n
.
2
0
1
9
,
p
p
.
6
9
–
7
4
.
d
o
i
:
1
0
.
1
1
0
9
/
S
D
S
.
2
0
1
9
.
0
0
0
-
5.
[
9
]
A
.
W
e
i
l
e
r
,
M
.
G
r
o
s
sn
i
k
l
a
u
s,
a
n
d
M
.
H
.
S
c
h
o
l
l
,
“
Ev
a
l
u
a
t
i
o
n
me
a
s
u
r
e
s
f
o
r
e
v
e
n
t
d
e
t
e
c
t
i
o
n
t
e
c
h
n
i
q
u
e
s
o
n
t
w
i
t
t
e
r
d
a
t
a
st
r
e
a
ms,
”
L
e
c
t
u
re
N
o
t
e
s
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
(
i
n
c
l
u
d
i
n
g
s
u
b
ser
i
e
s
L
e
c
t
u
r
e
N
o
t
e
s
i
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
L
e
c
t
u
re
N
o
t
e
s
i
n
Bi
o
i
n
f
o
rm
a
t
i
c
s)
,
v
o
l
.
9
1
4
7
,
p
p
.
1
0
8
–
1
1
9
,
2
0
1
5
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
3
1
9
-
2
0
4
2
4
-
6
_
1
1
.
[
1
0
]
A
.
W
e
i
l
e
r
,
M
.
G
r
o
ss
n
i
k
l
a
u
s
,
a
n
d
M
.
H
.
S
c
h
o
l
l
,
“
A
n
e
v
a
l
u
a
t
i
o
n
o
f
t
h
e
r
u
n
-
t
i
me
a
n
d
t
a
sk
-
b
a
s
e
d
p
e
r
f
o
r
m
a
n
c
e
o
f
e
v
e
n
t
d
e
t
e
c
t
i
o
n
t
e
c
h
n
i
q
u
e
s f
o
r
Tw
i
t
t
e
r
,
”
I
n
f
o
rm
a
t
i
o
n
S
y
s
t
e
m
s
,
v
o
l
.
6
2
,
p
p
.
2
0
7
–
2
1
9
,
D
e
c
.
2
0
1
6
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
s
.
2
0
1
6
.
0
1
.
0
0
3
.
[
1
1
]
E.
S
t
r
u
b
e
l
l
,
A
.
G
a
n
e
s
h
,
a
n
d
A
.
M
c
C
a
l
l
u
m
,
“
En
e
r
g
y
a
n
d
p
o
l
i
c
y
c
o
n
s
i
d
e
r
a
t
i
o
n
s
f
o
r
d
e
e
p
l
e
a
r
n
i
n
g
i
n
N
LP
,
”
A
C
L
2
0
1
9
-
5
7
t
h
A
n
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
re
n
c
e
,
p
p
.
3
6
4
5
–
3
6
5
0
,
2
0
2
0
.
[
1
2
]
J.
P
e
n
n
i
n
g
t
o
n
,
R
.
S
o
c
h
e
r
,
a
n
d
C
.
D
.
M
a
n
n
i
n
g
,
“
G
l
o
V
e
:
g
l
o
b
a
l
v
e
c
t
o
r
s
f
o
r
w
o
r
d
r
e
p
r
e
s
e
n
t
a
t
i
o
n
,
”
EM
N
L
P
2
0
1
4
-
2
0
1
4
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
i
n
g
,
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
r
e
n
c
e
,
p
p
.
1
5
3
2
–
1
5
4
3
,
2
0
1
4
,
d
o
i
:
1
0
.
3
1
1
5
/
v
1
/
d
1
4
-
1
1
6
2
.
[
1
3
]
T.
W
o
l
f
e
t
a
l
.
,
“
H
u
g
g
i
n
g
F
a
c
e
’
s
t
r
a
n
s
f
o
r
mers
:
st
a
t
e
-
of
-
t
h
e
-
a
r
t
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ss
i
n
g
,
”
a
r
Xi
v
p
re
p
r
i
n
t
1
9
1
0
.
0
3
7
7
1
,
p
p
.
1
–
8
,
2
0
1
9
.
[
1
4
]
G
.
H
i
n
t
o
n
,
O
.
V
i
n
y
a
l
s,
a
n
d
J.
D
e
a
n
,
“
D
i
st
i
l
l
i
n
g
t
h
e
k
n
o
w
l
e
d
g
e
i
n
a
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
a
rX
i
v
p
re
p
ri
n
t
1
5
0
3
.
0
2
5
3
1
,
M
a
r
.
2
0
1
5
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
1
5
0
3
.
0
2
5
3
1
.
[
1
5
]
A
.
F
.
A
d
o
m
a
,
N
.
M
.
H
e
n
r
y
,
a
n
d
W
.
C
h
e
n
,
“
C
o
mp
a
r
a
t
i
v
e
a
n
a
l
y
ses
o
f
b
e
r
t
,
r
o
b
e
r
t
a
,
d
i
s
t
i
l
b
e
r
t
,
a
n
d
x
l
n
e
t
f
o
r
t
e
x
t
-
b
a
se
d
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
,”
2
0
2
0
1
7
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
m
p
u
t
e
r
C
o
n
f
e
re
n
c
e
o
n
Wa
v
e
l
e
t
A
c
t
i
v
e
Me
d
i
a
T
e
c
h
n
o
l
o
g
y
a
n
d
I
n
f
o
rm
a
t
i
o
n
Pr
o
c
e
ss
i
n
g
,
I
C
C
W
AM
T
I
P
2
0
2
0
,
p
p
.
1
1
7
–
1
2
1
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
W
A
M
TI
P
5
1
6
1
2
.
2
0
2
0
.
9
3
1
7
3
7
9
.
[
1
6
]
D
.
C
h
a
t
t
e
r
j
e
e
,
“
M
a
k
i
n
g
n
e
u
r
a
l
m
a
c
h
i
n
e
r
e
a
d
i
n
g
c
o
m
p
r
e
h
e
n
si
o
n
f
a
st
e
r
,
”
a
r
Xi
v
p
re
p
ri
n
t
1
9
0
4
.
0
0
7
9
6
,
M
a
r
.
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
1
9
0
4
.
0
0
7
9
6
.
[
1
7
]
D
.
M
a
n
g
a
l
a
n
d
H
.
M
a
k
w
a
n
a
,
“
Ex
t
r
a
c
t
i
n
g
g
e
o
-
r
e
f
e
r
e
n
c
e
s
f
r
o
m
so
c
i
a
l
me
d
i
a
t
e
x
t
u
si
n
g
b
i
-
l
o
n
g
s
h
o
r
t
t
e
r
m
me
mo
r
y
n
e
t
w
o
r
k
s
,
”
I
n
d
o
n
e
si
a
n
J
o
u
r
n
a
l
o
f
El
e
c
t
r
i
c
a
l
E
n
g
i
n
e
e
r
i
n
g
a
n
d
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
3
5
,
n
o
.
2
,
p
p
.
1
2
6
3
–
1
2
7
0
,
A
u
g
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
e
e
c
s.
v
3
5
.
i
2
.
p
p
1
2
6
3
-
1
2
7
0
.
[
1
8
]
C
.
B
u
c
i
l
ǎ
,
R
.
C
a
r
u
a
n
a
,
a
n
d
A
.
N
i
c
u
l
e
sc
u
-
M
i
z
i
l
,
“
M
o
d
e
l
c
o
m
p
r
e
ssi
o
n
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
A
C
M
S
I
G
K
D
D
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
K
n
o
w
l
e
d
g
e
D
i
sc
o
v
e
ry
a
n
d
D
a
t
a
Mi
n
i
n
g
,
N
e
w
Y
o
r
k
,
N
Y
,
U
S
A
:
A
C
M
,
A
u
g
.
2
0
0
6
,
p
p
.
5
3
5
–
5
4
1
.
d
o
i
:
1
0
.
1
1
4
5
/
1
1
5
0
4
0
2
.
1
1
5
0
4
6
4
.
[
1
9
]
A
.
L
.
M
a
a
s,
R
.
E.
D
a
l
y
,
P
.
T
.
P
h
a
m,
D
.
H
u
a
n
g
,
A
.
Y
.
N
g
,
a
n
d
C
.
P
o
t
t
s,
“
L
e
a
r
n
i
n
g
w
o
r
d
v
e
c
t
o
r
s
f
o
r
se
n
t
i
me
n
t
a
n
a
l
y
si
s,
”
AC
L
-
H
L
T
2
0
1
1
-
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
4
9
t
h
An
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
Ass
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s:
H
u
m
a
n
L
a
n
g
u
a
g
e
T
e
c
h
n
o
l
o
g
i
e
s
,
v
o
l
.
1
,
p
p
.
1
4
2
–
1
5
0
,
2
0
1
1
.
[
2
0
]
A
.
Tu
r
c
h
i
n
,
S
.
M
a
s
h
a
r
sk
y
,
a
n
d
M
.
Z
i
t
n
i
k
,
“
C
o
mp
a
r
i
s
o
n
o
f
B
E
R
T
i
mp
l
e
m
e
n
t
a
t
i
o
n
s
f
o
r
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
s
si
n
g
o
f
n
a
r
r
a
t
i
v
e
med
i
c
a
l
d
o
c
u
me
n
t
s
,
”
I
n
f
o
rm
a
t
i
c
s
i
n
Me
d
i
c
i
n
e
U
n
l
o
c
k
e
d
,
v
o
l
.
3
6
,
p
.
1
0
1
1
3
9
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
m
u
.
2
0
2
2
.
1
0
1
1
3
9
.
[
2
1
]
R
.
T
a
n
g
,
Y
.
L
u
,
L
.
Li
u
,
L.
M
o
u
,
O
.
V
e
c
h
t
o
mo
v
a
,
a
n
d
J.
L
i
n
,
“
D
i
st
i
l
l
i
n
g
t
a
sk
-
s
p
e
c
i
f
i
c
k
n
o
w
l
e
d
g
e
f
r
o
m
B
ER
T
i
n
t
o
si
mp
l
e
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
M
a
r
.
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
1
9
0
3
.
1
2
1
3
6
[
2
2
]
S
.
S
u
b
r
a
ma
n
i
a
n
,
A
.
R
a
h
i
m
i
,
T
.
B
a
l
d
w
i
n
,
T
.
C
o
h
n
,
a
n
d
L.
F
r
e
r
ma
n
n
,
“
F
a
i
r
n
e
ss
-
a
w
a
r
e
c
l
a
ss
i
m
b
a
l
a
n
c
e
d
l
e
a
r
n
i
n
g
,
”
EM
N
L
P
2
0
2
1
-
2
0
2
1
C
o
n
f
e
r
e
n
c
e
o
n
Em
p
i
ri
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
,
Pr
o
c
e
e
d
i
n
g
s
,
p
p
.
2
0
4
5
–
2
0
5
1
,
2
0
2
1
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
1
.
e
mn
l
p
-
mai
n
.
1
5
5
.
[
2
3
]
N
.
Ze
i
n
a
l
i
,
A
.
A
l
B
a
s
h
a
y
r
e
h
,
W
.
F
a
n
,
a
n
d
S
.
G
.
W
h
i
t
e
,
“
C
o
mp
a
r
i
s
o
n
o
f
B
ER
T
I
mp
l
e
me
n
t
a
t
i
o
n
s
f
o
r
e
n
h
a
n
c
e
d
c
a
n
c
e
r
s
y
m
p
t
o
ms
e
x
t
r
a
c
t
i
o
n
f
r
o
m
e
l
e
c
t
r
o
n
i
c
h
e
a
l
t
h
r
e
c
o
r
d
s,”
i
n
Pr
o
c
e
e
d
i
n
g
s
-
2
0
2
4
I
E
EE
1
s
t
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
f
o
r
Me
d
i
c
i
n
e
,
H
e
a
l
t
h
a
n
d
C
a
re
,
AI
MHC
2
0
2
4
,
I
EEE,
F
e
b
.
2
0
2
4
,
p
p
.
1
8
–
1
9
.
d
o
i
:
1
0
.
1
1
0
9
/
A
I
M
H
C
5
9
8
1
1
.
2
0
2
4
.
0
0
0
1
1
.
[
2
4
]
D
.
C
o
r
t
i
z
,
“
Ex
p
l
o
r
i
n
g
t
r
a
n
sf
o
r
m
e
r
s
mo
d
e
l
s
f
o
r
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
:
a
c
o
mp
a
r
i
s
i
o
n
o
f
B
E
R
T
,
D
i
s
t
i
l
B
ER
T,
R
o
B
E
R
T
a
,
X
LN
ET
a
n
d
ELE
C
TR
A
,
”
i
n
AC
M
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
Pr
o
c
e
e
d
i
n
g
S
e
ri
e
s
,
N
e
w
Y
o
r
k
,
N
Y
,
U
S
A
:
A
C
M
,
A
u
g
.
2
0
2
2
,
p
p
.
2
3
0
–
2
3
4
.
d
o
i
:
1
0
.
1
1
4
5
/
3
5
6
2
0
0
7
.
3
5
6
2
0
5
1
.
[
2
5
]
I
.
B
e
l
t
a
g
y
,
K
.
Lo
,
a
n
d
A
.
C
o
h
a
n
,
“
S
c
i
B
ER
T
:
a
p
r
e
t
r
a
i
n
e
d
l
a
n
g
u
a
g
e
m
o
d
e
l
f
o
r
sci
e
n
t
i
f
i
c
t
e
x
t
,
”
a
r
Xi
v
p
re
p
ri
n
t
1
9
0
3
.
1
0
6
7
6
,
M
a
r
.
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
1
9
0
3
.
1
0
6
7
6
.
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.
3
8
,
No
.
1
,
Ap
r
il
20
2
5
:
4
3
9
-
4
4
6
446
[
2
6
]
J.
-
S
.
Le
e
a
n
d
J
.
H
si
a
n
g
,
“
P
a
t
e
n
t
B
E
R
T:
p
a
t
e
n
t
c
l
a
ss
i
f
i
c
a
t
i
o
n
w
i
t
h
f
i
n
e
-
t
u
n
i
n
g
a
p
r
e
-
t
r
a
i
n
e
d
B
E
R
T
mo
d
e
l
,”
a
rX
i
v
p
r
e
p
r
i
n
t
1
9
0
6
.
0
2
1
2
4
,
M
a
y
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
1
9
0
6
.
0
2
1
2
4
[
2
7
]
I
.
C
h
a
l
k
i
d
i
s
,
M
.
F
e
r
g
a
d
i
o
t
i
s
,
P
.
M
a
l
a
k
a
s
i
o
t
i
s
,
N
.
A
l
e
t
r
a
s,
a
n
d
I
.
A
n
d
r
o
u
t
s
o
p
o
u
l
o
s
,
“
LEG
A
L
-
B
ER
T:
T
h
e
m
u
p
p
e
t
s
s
t
r
a
i
g
h
t
o
u
t
o
f
l
a
w
sc
h
o
o
l
,
”
Fi
n
d
i
n
g
s
o
f
t
h
e
Ass
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s
F
i
n
d
i
n
g
s
o
f
AC
L
:
E
MN
L
P
2
0
2
0
,
p
p
.
2
8
9
8
–
2
9
0
4
,
O
c
t
.
2
0
2
0
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
f
i
n
d
i
n
g
s
-
e
mn
l
p
.
2
6
1
.
[
2
8
]
M
.
B
.
A
.
M
c
D
e
r
m
o
t
t
,
B
.
Y
a
p
,
P
.
S
z
o
l
o
v
i
t
s
,
a
n
d
M
.
Z
i
t
n
i
k
,
“
S
t
r
u
c
t
u
r
e
i
n
d
u
c
i
n
g
p
r
e
-
t
r
a
i
n
i
n
g
,
”
a
r
Xi
v
p
re
p
ri
n
t
2
1
0
3
.
1
0
3
3
4
,
M
a
r
.
2
0
2
1
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
2
1
0
3
.
1
0
3
3
4
[
2
9
]
X
.
D
u
a
n
d
C
.
C
a
r
d
i
e
,
“
E
v
e
n
t
e
x
t
r
a
c
t
i
o
n
b
y
a
n
sw
e
r
i
n
g
(
a
l
m
o
st
)
n
a
t
u
r
a
l
q
u
e
st
i
o
n
s
,
”
i
n
EM
N
L
P
2
0
2
0
-
2
0
2
0
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
M
e
t
h
o
d
s
i
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
,
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
r
e
n
c
e
,
S
t
r
o
u
d
s
b
u
r
g
,
P
A
,
U
S
A
:
A
ss
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s,
2
0
2
0
,
p
p
.
6
7
1
–
6
8
3
.
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
e
mn
l
p
-
ma
i
n
.
4
9
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Dha
r
m
e
n
d
r
a
Ma
n
g
a
l
is
a
ss
istan
t
p
r
o
fe
ss
o
r
in
M
e
d
i
-
Ca
p
s
Un
i
v
e
rsity
,
I
n
d
o
re
a
lso
P
h
.
D
.
sc
h
o
lar
a
t
IET
,
DAVV,
In
d
o
re
.
T
h
e
re
se
a
rc
h
in
tere
st
in
c
l
u
d
e
s
NLP
a
n
d
m
a
c
h
in
e
lea
rn
in
g
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
m
a
n
g
a
ld
h
a
rm
e
n
d
ra
8
3
@g
m
a
il
.
c
o
m
.
Dr
.
H
e
m
a
n
t
M
a
k
wa
n
a
is
a
ss
o
c
iate
p
ro
fe
ss
o
r
a
t
IE
T,
DAVV,
I
n
d
o
re
.
T
h
e
re
se
a
rc
h
in
tere
st
i
n
c
lu
d
e
s
c
o
m
p
u
ter
g
ra
p
h
ics
a
n
d
c
o
m
p
u
ter
a
rc
h
it
e
c
tu
re
.
He
is
P
h
.
D
.
in
c
o
m
p
u
ter
e
n
g
in
e
e
rin
g
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
h
m
a
k
wa
n
a
@ie
td
a
v
v
.
e
d
u
.
in
.
Evaluation Warning : The document was created with Spire.PDF for Python.