I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
3
,
J
une
2025
, pp.
2367
~
2378
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
3
.pp
2367
-
2378
2367
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
M
e
d
i
c
al
B
E
R
T
:
e
n
h
an
c
i
n
g b
i
om
e
d
i
c
al
n
at
u
r
al
l
an
g
u
age
p
r
oc
e
ss
i
n
g u
si
n
g
p
r
e
t
r
ai
n
e
d
B
E
R
T
-
b
ase
d
m
od
e
l
K
. S
ah
it
R
e
d
d
y, N
. R
agave
n
d
e
r
an
, V
as
an
t
h
K
., G
a
n
e
s
h
N
. N
ai
k
, V
is
h
al
ak
s
h
i
P
r
ab
h
u
,
N
agar
aj
a G
.
S
.
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
, R
. V
.
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
ng, B
e
nga
l
ur
u, I
ndi
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
r
10
,
2024
R
e
vi
s
e
d
F
e
b
28
,
2025
A
c
c
e
pt
e
d
M
a
r
15
,
2025
Recent
advances
in
natural
language
processing
(NLP)
have
been
dri
ven
by
pretrained
language
models
like
BERT,
RoBERTa,
T5,
and
GPT.
These
models
excel
at
understanding
complex
texts,
but
biomedical
literatur
e,
with
its
domain
-
specific
terminology,
poses
challenges
that
models
like
Word2Vec
and
bidirecti
onal
long
short
-
term
memory
(Bi
-
LSTM)
can'
t
fully
address.
GPT
and
T5,
d
espite
capturing
context,
fall
short
in
tasks
n
eeding
bidirectional
understanding,
unlike
BERT.
Addressing
this,
we
pr
oposed
MedicalBER
T,
a
pretrained
BERT
model
trained
on
a
large
bio
medical
dataset
and
equipped
with
domain
-
specific
vocabulary
that
enhanc
es
the
comprehens
ion
of
biomedi
cal
terminol
ogy.
MedicalB
ERT
model
is
further
optimized
and
fine
-
tuned
to
address
diverse
t
asks,
including
named
entity
recognition
,
relation
extraction,
question
answering,
sentence
similari
t
y,
and
document
classification.
Performance
metrics
such
as
the
F1
-
score,
accuracy,
and
Pearson
correlatio
n
are
employed
to
showcase
the
effi
ciency
of
our
model
in
comp
arison
to
other
BERT
-
based
models
such
as
Bio
BERT,
SciBERT
,
and
ClinicalBERT.
MedicalBERT
outperforms
these
models
on
most
of
the
benchmarks,
and
surpasses
the
general
-
purpose
BERT
mo
del
by
5.67%
on
average
across
all
the
tasks
evaluated
respectively.
This
wo
rk
also
underscores
the
potential
of
leveraging
pretrained
BERT
models
for
medical
NLP tasks, demo
nstrating the
effectiveness of
transfer learning techni
ques in
capturing
domain
-
specific infor
mation.
K
e
y
w
o
r
d
s
:
B
E
R
T
N
a
m
e
d
e
nt
it
y r
e
c
ogni
ti
on
N
a
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
P
r
e
-
tr
a
in
e
d m
ode
ls
R
e
la
ti
on
e
xt
r
a
c
ti
on
T
r
a
ns
f
or
m
e
r
s
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
N
.
R
a
ga
ve
nde
r
a
n
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
, R
.
V
.
C
ol
le
g
e
of
E
ngi
ne
e
r
in
g
M
ys
or
e
R
d, R
V
V
id
ya
ni
ke
ta
n P
os
t
,
B
e
nga
lu
r
u, I
ndi
a
E
m
a
il
:
nr
a
ga
ve
nde
r
a
n.c
s
22@
r
vc
e
.e
du.i
n
1.
I
N
T
R
O
D
U
C
T
I
O
N
I
n
th
e
r
e
a
lm
of
na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
(
N
L
P
)
,
th
e
r
e
vol
ut
io
na
r
y
in
f
lu
e
nc
e
of
s
ophi
s
ti
c
a
te
d
tr
a
ns
f
or
m
e
r
m
ode
ls
is
obs
e
r
va
bl
e
a
c
r
os
s
m
ul
ti
pl
e
f
ie
ld
s
.
I
n
th
e
he
a
lt
hc
a
r
e
in
dus
tr
y,
e
le
c
tr
oni
c
he
a
lt
h
r
e
c
or
ds
(
E
H
R
s
)
a
r
e
a
s
ub
s
ta
nt
ia
l
s
our
c
e
of
va
lu
a
bl
e
da
t
a
,
a
nd
B
E
R
T
ha
s
pr
ove
n
to
be
a
pow
e
r
f
ul
to
ol
f
or
de
c
ip
he
r
in
g
th
e
i
nt
r
ic
a
te
c
ont
e
xt
ua
l
de
ta
il
s
e
m
be
dd
e
d w
it
hi
n uns
tr
uc
tu
r
e
d m
e
di
c
a
l
te
xt
.
B
E
R
T
[
1]
,
s
hor
t
f
or
bi
di
r
e
c
ti
ona
l
e
nc
ode
r
r
e
p
r
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
,
is
a
f
o
r
m
id
a
bl
e
N
L
P
m
ode
l
th
a
t
w
a
s
in
tr
oduc
e
d
by
G
oogl
e
in
2018.
I
ts
k
e
y
in
nov
a
ti
on
li
e
s
in
it
s
bi
di
r
e
c
ti
ona
l
a
ppr
oa
c
h,
w
hi
c
h
e
na
bl
e
s
it
to
ta
ke
in
to
a
c
c
ount
th
e
c
ont
e
xt
of
w
or
ds
f
r
om
bo
th
di
r
e
c
ti
ons
.
P
r
e
tr
a
in
e
d
on
e
xt
e
ns
iv
e
te
xt
da
ta
,
B
E
R
T
e
xc
e
ls
in
c
a
pt
ur
in
g
nua
nc
e
d
la
ngua
g
e
r
e
la
ti
ons
hi
ps
,
m
a
ki
ng
it
a
c
or
ne
r
s
to
ne
in
va
r
io
us
N
L
P
a
ppl
ic
a
ti
ons
.
I
n
th
e
le
g
a
l
dom
a
in
,
B
E
R
T
a
s
s
i
s
ts
in
p
a
r
s
in
g
a
nd
e
xt
r
a
c
ti
ng
r
e
le
va
nt
in
f
or
m
a
ti
on
f
r
om
le
ga
l
doc
um
e
nt
s
.
W
hi
le
B
E
R
T
ha
s
de
m
ons
tr
a
te
d
pr
of
ic
ie
nc
y
in
ge
ne
r
a
l
la
ngua
ge
unde
r
s
ta
ndi
ng,
th
e
he
a
lt
hc
a
r
e
dom
a
in
de
m
a
nds
a
s
pe
c
ia
li
z
e
d
a
ppr
oa
c
h.
C
onve
nt
io
na
l
N
L
P
te
c
hni
que
s
,
li
ke
b
a
g
-
of
-
w
or
ds
(
B
oW
)
[
2]
a
nd
n
-
gr
a
m
s
[
3]
,
s
tr
uggl
e
w
it
h
c
ont
e
xt
ua
l
unde
r
s
ta
ndi
ng,
tr
e
a
ti
ng
w
or
ds
in
de
pe
nde
nt
ly
a
nd
of
t
e
n
le
a
di
ng
to
in
f
or
m
a
ti
on
lo
s
s
, e
s
pe
c
ia
ll
y
w
it
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2367
-
2378
2368
va
r
yi
ng
te
xt
le
ngt
hs
.
M
e
th
ods
li
ke
T
F
-
I
D
F
a
ls
o
f
a
c
e
is
s
ue
s
w
i
th
pol
ys
e
m
y,
m
a
nua
l
f
e
a
tu
r
e
e
ngi
ne
e
r
in
g,
a
nd
c
a
pt
ur
in
g
lo
ng
-
r
a
nge
de
pe
nde
nc
ie
s
,
li
m
it
in
g
pe
r
f
or
m
a
nc
e
a
n
d
tr
a
ns
f
e
r
le
a
r
ni
ng.
T
r
a
ns
f
or
m
e
r
m
ode
ls
li
ke
B
E
R
T
s
ol
ve
th
e
s
e
is
s
ue
s
by
u
s
in
g
de
e
p
bi
di
r
e
c
ti
ona
l
c
ont
e
xt
u
a
l
r
e
pr
e
s
e
nt
a
ti
ons
,
im
pr
ovi
ng
a
da
pt
a
bi
li
ty
a
nd
a
c
c
ur
a
c
y a
c
r
o
s
s
di
ve
r
s
e
N
L
P
t
a
s
ks
.
T
hi
s
w
or
k
f
oc
u
s
e
s
on
pr
opos
in
g
a
B
E
R
T
-
ba
s
e
d
m
ode
l,
M
e
di
c
a
lB
E
R
T
,
f
or
bi
om
e
di
c
a
l
te
xt
m
in
in
g
a
nd
N
L
P
,
w
it
h
s
pe
c
ia
l
a
tt
e
nt
io
n
gi
ve
n
to
th
e
uni
que
c
ha
r
a
c
te
r
is
ti
c
s
th
a
t
di
f
f
e
r
e
nt
ia
te
th
is
m
ode
l
f
r
om
ot
he
r
s
pe
c
ia
li
z
e
d
B
E
R
T
m
ode
ls
in
th
is
dom
a
in
.
T
hr
ough
e
xpe
r
im
e
nt
a
ti
on
a
nd
e
va
lu
a
ti
on
by
pr
e
-
tr
a
in
in
g
a
nd
f
in
e
-
tu
ni
ng,
w
e
de
m
ons
tr
a
te
how
th
is
pr
opos
e
d
m
ode
l
s
ta
nds
out
,
by
c
om
pa
r
in
g
th
e
out
c
om
e
of
our
m
ode
l
,
w
it
h
ot
he
r
B
E
R
T
-
ba
s
e
d
m
ode
ls
th
a
t
onc
e
e
m
e
r
ge
d
a
s
s
ta
te
-
of
-
th
e
-
a
r
t
m
ode
ls
a
nd
c
ont
r
ib
ut
e
d
to
th
e
B
io
N
L
P
c
om
m
uni
ty
.
S
e
c
ti
on
5
pr
e
s
e
nt
s
a
c
om
pa
r
a
ti
ve
a
na
ly
s
is
a
nd
r
e
s
ul
ts
of
our
pr
opos
e
d
m
ode
l
w
i
th
th
os
e
of
ot
he
r
m
ode
ls
. P
r
e
tr
a
in
in
g r
e
f
e
r
s
t
o t
he
i
ni
ti
a
l
pha
s
e
w
he
r
e
t
he
m
ode
l
le
a
r
ns
l
a
ngua
ge
pa
tt
e
r
n
s
f
r
om
a
l
a
r
ge
c
or
pus
of
te
xt
da
ta
,
c
a
pt
ur
in
g
ge
ne
r
a
l
la
ngua
ge
f
e
a
tu
r
e
s
.
F
in
e
-
tu
ni
ng,
on
th
e
ot
he
r
ha
nd,
in
vol
ve
s
f
ur
th
e
r
t
r
a
in
in
g
th
e
pr
e
tr
a
in
e
d
m
ode
l
on
a
s
pe
c
if
ic
ta
s
k
or
dom
a
in
to
e
nha
nc
e
it
s
pe
r
f
or
m
a
nc
e
f
or
th
a
t
pa
r
ti
c
ul
a
r
a
ppl
ic
a
ti
on.
D
e
ta
il
e
d
e
va
lu
a
ti
on
m
e
tr
ic
s
il
lu
s
tr
a
te
th
e
m
ode
l’
s
s
tr
e
ngt
hs
a
nd
hi
ghl
ig
ht
a
r
e
a
s
f
or
pot
e
nt
ia
l
im
pr
ove
m
e
nt
.
T
hi
s
be
nc
hm
a
r
ki
ng not onl
y de
m
ons
tr
a
te
s
our
pr
opos
e
d m
ode
l'
s
i
m
pr
ove
m
e
nt
s
but
a
ls
o pr
ovi
de
s
i
ns
ig
ht
s
i
nt
o
th
e
s
pe
c
if
ic
s
tr
e
ngt
hs
a
nd
w
e
a
kne
s
s
e
s
r
e
la
ti
ve
to
e
xi
s
ti
ng
m
ode
ls
.
T
hi
s
w
or
k
a
ls
o
hi
ghl
ig
ht
s
th
e
s
ig
ni
f
ic
a
nc
e
of
us
in
g c
us
to
m
t
a
il
or
e
d voc
a
bul
a
r
y a
nd a
l
a
r
ge
r
t
r
a
in
in
g s
e
t
a
n
d i
ts
i
m
pa
c
t
on t
he
pe
r
f
or
m
a
nc
e
of
t
he
m
ode
l.
T
h
is
p
a
p
e
r
c
o
ns
is
t
s
of
6
s
e
c
ti
o
ns
–
1
st
s
e
c
t
io
n
i
s
t
he
i
n
tr
od
uc
ti
o
n,
f
ol
lo
w
e
d
by t
he
2
nd
s
e
c
ti
o
n w
hi
c
h
gi
v
e
s
a
br
i
e
f
kn
ow
l
e
d
g
e
a
b
out
t
he
r
e
l
a
te
d
B
E
R
T
-
b
a
s
e
d
m
o
de
ls
t
h
a
t
w
i
ll
b
e
us
e
d
f
or
pe
r
f
or
m
a
n
c
e
c
o
m
p
a
r
i
s
o
n
pr
e
s
e
nt
e
d
in
th
e
r
e
s
ul
t
s
s
e
c
ti
on
,
i
.e
., 5
th
s
e
c
ti
on.
T
h
e
3
rd
s
e
c
t
io
n
g
iv
e
s
a
n
o
ve
r
vi
e
w
of
B
E
R
T
a
r
c
hi
t
e
c
tu
r
e
, a
nd a
ll
th
e
be
nc
hm
a
r
k
da
t
a
s
e
t
s
a
nd
t
he
c
or
r
e
s
p
on
di
n
g
N
L
P
ta
s
k
s
pe
r
f
or
m
e
d
o
n
e
a
c
h.
C
o
nt
i
nui
ng
f
or
w
a
r
d
,
th
e
4
th
s
e
c
t
io
n
p
r
ov
id
e
s
t
he
m
e
t
hod
i
n
w
h
ic
h
M
e
di
c
a
l
B
E
R
T
w
a
s
tr
a
i
ne
d
a
nd
f
i
ne
-
tu
ne
d
to
gi
ve
t
he
f
i
na
l
p
r
op
o
s
e
d
m
od
e
l,
w
h
ic
h
ou
tp
e
r
f
or
m
s
m
o
s
t
o
f
th
e
s
ta
te
-
of
-
t
he
-
a
r
t
m
od
e
l
s
i
n
th
e
bi
o
m
e
di
c
a
l
dom
a
i
n.
F
in
a
ll
y,
t
he
5
th
s
e
c
t
io
n
pr
ov
id
e
s
t
he
r
e
s
ul
t
s
of
our
w
o
r
k
a
n
d f
ur
th
e
r
di
s
c
u
s
s
io
n
a
b
out
t
h
e
w
or
k
, f
ol
l
ow
e
d
by
t
h
e
c
on
c
lu
s
i
on
s
e
c
ti
on
.
2.
R
E
L
A
T
E
D
M
O
D
E
L
S
A
N
D
WORKS
2.1. B
io
B
E
R
T
B
io
B
E
R
T
[
4]
,
a
dom
a
in
-
s
pe
c
if
ic
li
ngui
s
ti
c
por
tr
a
ya
l
f
r
a
m
e
w
or
k,
e
m
e
r
ge
s
f
r
om
a
n
in
it
ia
l
B
E
R
T
c
he
c
kpoi
nt
tr
a
in
e
d
on
bi
om
e
di
c
a
l
t
e
xt
f
r
om
P
ubM
e
d
A
b
s
tr
a
c
ts
a
nd
P
ubM
e
d
C
e
nt
r
a
l.
I
ts
a
da
pt
a
ti
on
f
or
bi
om
e
di
c
a
l
te
xt
m
in
in
g
a
c
knowle
dge
s
th
e
gr
ow
in
g
im
por
ta
nc
e
of
th
is
f
ie
ld
,
dr
iv
e
n
by
th
e
r
a
pi
d
e
xpa
n
s
io
n
of
bi
om
e
di
c
a
l
doc
um
e
nt
vol
um
e
s
.
N
L
P
a
dva
nc
e
m
e
nt
s
a
ppl
ie
d
d
ir
e
c
tl
y
to
bi
om
e
di
c
a
l
te
xt
m
in
in
g
of
te
n
yi
e
ld
s
ubopti
m
a
l
out
c
om
e
s
due
to
di
s
tr
ib
ut
io
n
s
hi
f
ts
f
r
om
ge
ne
r
a
l
dom
a
in
c
or
por
a
to
bi
om
e
di
c
a
l
c
or
por
a
.
B
io
B
E
R
T
out
pe
r
f
or
m
s
B
E
R
T
a
nd e
a
r
li
e
r
t
op
-
pe
r
f
or
m
in
g m
ode
l
s
a
c
r
os
s
v
a
r
io
us
bi
om
e
di
c
a
l
te
xt
m
in
in
g t
a
s
ks
,
in
c
lu
di
ng
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on,
r
e
la
ti
on
e
xt
r
a
c
ti
on,
a
nd
que
s
ti
on
a
ns
w
e
r
in
g.
A
lt
hough
e
f
f
e
c
ti
ve
in
ge
ne
r
a
l
bi
om
e
di
c
a
l
N
L
P
ta
s
ks
,
B
io
B
E
R
T
’
s
r
e
li
a
nc
e
on
a
m
or
e
ge
ne
r
a
li
z
e
d
voc
a
bul
a
r
y
oc
c
a
s
io
na
ll
y
r
e
duc
e
s
it
s
pe
r
f
or
m
a
nc
e
on
da
t
a
s
e
t
s
w
it
h
hi
ghl
y
s
pe
c
if
ic
te
r
m
in
ol
og
y.
I
n
c
ont
r
a
s
t,
M
e
di
c
a
lB
E
R
T
u
s
e
s
a
c
u
s
to
m
bi
om
e
di
c
a
l
voc
a
bul
a
r
y
a
li
gne
d
w
it
h
B
E
R
T
to
ke
ni
z
a
ti
on,
a
ll
ow
in
g
it
to
be
tt
e
r
unde
r
s
ta
nd
a
nd
in
te
r
pr
e
t
c
om
pl
e
x t
e
r
m
s
uni
que
t
o bi
om
e
di
c
a
l
li
te
r
a
tu
r
e
.
2.2. B
lu
e
B
E
R
T
I
ns
pi
r
e
d
by
th
e
s
uc
c
e
s
s
of
th
e
ge
ne
r
a
l
la
ngua
ge
unde
r
s
ta
ndi
ng
e
va
lu
a
ti
on
be
nc
hm
a
r
k,
th
is
w
or
k
in
tr
oduc
e
d
th
e
bi
om
e
di
c
a
l
la
ngua
g
e
unde
r
s
ta
ndi
ng
e
va
lu
a
ti
on
(
B
L
U
E
)
be
nc
hm
a
r
k
[
5]
.
E
va
lu
a
ti
ons
on
m
ul
ti
pl
e
ba
s
e
li
ne
s
i
nvol
vi
ng B
E
R
T
[
1]
a
nd E
L
M
o
[
6]
pr
ovi
de
a
c
om
pa
r
a
ti
ve
a
na
ly
s
is
of
pr
e
tr
a
in
in
g
l
a
ngua
ge
r
e
pr
e
s
e
nt
a
ti
ons
in
th
e
bi
om
e
di
c
a
l
f
ie
ld
.
T
he
B
E
R
T
m
ode
l
pr
e
tr
a
in
e
d
on
P
ubM
e
d
a
bs
tr
a
c
ts
a
nd
M
I
M
I
C
-
III
c
li
ni
c
a
l
not
e
s
a
c
hi
e
ve
d
th
e
b
e
s
t
r
e
s
ul
ts
a
m
ong
th
e
e
v
a
lu
a
te
d
ba
s
e
li
ne
s
,
e
xt
e
ndi
ng
b
e
nc
hm
a
r
ki
ng
s
u
c
c
e
s
s
to
bi
om
e
di
c
in
e
.
T
he
B
L
U
E
be
nc
hm
a
r
k
c
om
pr
is
e
s
f
iv
e
a
s
s
ig
nm
e
nt
s
w
it
h
te
n
da
ta
s
e
ts
,
e
nc
om
pa
s
s
in
g
a
va
r
ie
ty
of
bi
om
e
di
c
a
l
a
nd
c
li
ni
c
a
l
do
c
um
e
nt
s
,
in
c
lu
di
ng
v
a
r
io
us
da
ta
s
e
t
s
iz
e
s
a
nd
di
f
f
ic
ul
ti
e
s
.
B
y
f
oc
us
in
g
n
a
r
r
ow
ly
on
P
ubM
e
d
a
nd
M
I
M
I
C
-
I
I
I
da
ta
,
B
lu
e
B
E
R
T
m
a
y
c
a
pt
ur
e
pa
tt
e
r
ns
a
nd
te
r
m
in
ol
ogy
s
p
e
c
if
ic
to
th
e
s
e
da
t
a
s
e
t
s
,
pot
e
nt
ia
ll
y
li
m
it
in
g
i
ts
ge
ne
r
a
li
z
a
bi
li
ty
a
c
r
os
s
br
oa
de
r
bi
om
e
di
c
a
l
or
c
li
ni
c
a
l
te
xt
s
.
T
hi
s
c
a
n
le
a
d
to
r
e
duc
e
d
pe
r
f
or
m
a
nc
e
on da
ta
s
e
ts
t
ha
t
us
e
di
ve
r
s
e
or
l
e
s
s
-
c
om
m
on me
di
c
a
l
voc
a
bul
a
r
y, a
s
t
he
m
ode
l
m
a
y ha
ve
l
e
a
r
ne
d
pa
tt
e
r
ns
th
a
t
a
r
e
ove
r
ly
s
pe
c
if
ic
to
it
s
pr
e
tr
a
in
in
g
d
a
ta
.
I
n
c
om
pa
r
is
on,
M
e
di
c
a
lB
E
R
T
’
s
pr
e
tr
a
in
in
g
on
e
xt
e
ns
iv
e
bi
om
e
di
c
a
l
da
ta
a
ddr
e
s
s
e
s
th
e
s
e
is
s
u
e
s
,
s
how
in
g
r
obus
t
pe
r
f
or
m
a
nc
e
a
c
r
os
s
va
r
ie
d
bi
om
e
di
c
a
l
be
nc
hm
a
r
ks
w
it
hout
t
he
s
a
m
e
ov
e
r
f
it
ti
ng l
im
it
a
ti
ons
.
2.3. Clin
ic
al
B
E
R
T
C
li
ni
c
a
lB
E
R
T
[
7]
,
pr
e
tr
a
in
e
d
on
c
li
ni
c
a
l
not
e
s
,
un
c
ove
r
s
hi
gh
-
qua
li
ty
r
e
la
ti
ons
hi
ps
be
twe
e
n
m
e
di
c
a
l
c
onc
e
pt
s
,
m
a
xi
m
iz
in
g
th
e
us
e
of
c
li
ni
c
a
l
doc
um
e
nt
a
ti
on.
I
t
de
ve
lo
ps
a
nd
e
va
lu
a
te
s
a
c
ont
in
uous
r
e
pr
e
s
e
nt
a
ti
on
of
hi
gh
-
di
m
e
ns
io
na
l
a
nd
s
p
a
r
s
e
c
li
ni
c
a
l
not
e
s
,
e
s
ti
m
a
ti
ng
th
e
li
ke
li
hood
of
pa
ti
e
nt
s
be
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
e
di
c
al
B
E
R
T
:
e
nhanc
in
g bi
om
e
di
c
al
nat
ur
al
l
anguage
pr
oc
e
s
s
in
g us
in
g pr
e
tr
ai
ne
d
…
(
K
. Sahit
R
e
ddy
)
2369
r
e
a
dm
it
te
d
to
th
e
hos
pi
ta
l
w
it
hi
n
30
da
ys
a
t
di
f
f
e
r
e
nt
in
te
r
v
a
ls
,
le
ve
r
a
gi
ng
th
e
B
E
R
T
m
ode
l.
S
ur
pa
s
s
in
g
m
ul
ti
pl
e
ba
s
e
li
ne
s
in
f
or
e
c
a
s
ti
ng
hos
pi
ta
l
r
e
a
dm
is
s
io
ns
w
it
hi
n
a
30
-
da
y
ti
m
e
f
r
a
m
e
,
ut
il
iz
in
g
di
s
c
ha
r
ge
s
um
m
a
r
ie
s
a
nd
pr
e
li
m
in
a
r
y
not
e
s
f
r
om
th
e
in
te
ns
iv
e
c
a
r
e
uni
t,
C
li
ni
c
a
lB
E
R
T
'
s
e
f
f
e
c
ti
ve
ne
s
s
in
r
e
pr
e
s
e
nt
in
g
c
li
ni
c
a
l
not
e
s
m
a
y
v
a
r
y
a
c
r
os
s
di
f
f
e
r
e
nt
c
li
ni
c
a
l
dom
a
in
s
,
r
e
qui
r
in
g
f
ur
th
e
r
va
li
da
ti
on.
H
ow
e
ve
r
,
it
s
pe
r
f
or
m
a
nc
e
va
r
ie
s
a
c
r
os
s
c
li
ni
c
a
l
dom
a
in
s
due
to
th
e
di
ve
r
s
it
y
in
m
e
di
c
a
l
la
ngua
ge
a
nd
te
r
m
in
ol
ogy
us
e
d
in
di
f
f
e
r
e
nt
s
pe
c
ia
lt
ie
s
.
F
or
in
s
ta
nc
e
,
w
hi
le
it
e
xc
e
ls
w
it
h
ge
ne
r
a
l
pa
ti
e
nt
da
ta
,
ta
s
ks
in
vol
vi
ng
ni
c
he
m
e
di
c
a
l
f
ie
ld
s
or
r
a
r
e
c
li
ni
c
a
l
c
ondi
ti
ons
m
a
y
s
how
r
e
duc
e
d
a
c
c
ur
a
c
y.
M
or
e
e
xt
e
ns
iv
e
va
li
da
ti
on
on
s
pe
c
ia
li
z
e
d
da
ta
s
e
ts
w
oul
d
he
lp
a
s
s
e
s
s
C
li
ni
c
a
lB
E
R
T
'
s
a
da
pt
a
bi
li
ty
a
c
r
os
s
va
r
io
us
c
li
ni
c
a
l
s
ub
-
dom
a
in
s
,
hi
ghl
ig
ht
in
g
w
he
r
e
M
e
di
c
a
lB
E
R
T
’
s
br
oa
d
bi
om
e
di
c
a
l
pr
e
tr
a
in
in
g
of
f
e
r
s
a
m
or
e
r
obus
t
a
lt
e
r
na
ti
ve
f
or
di
ve
r
s
e
c
li
ni
c
a
l
a
nd
bi
om
e
di
c
a
l
a
ppl
ic
a
ti
ons
.
2.4. S
c
iB
E
R
T
G
iv
e
n
th
e
s
c
a
r
c
it
y
of
e
xt
e
ns
iv
e
ly
a
nnot
a
te
d
d
a
ta
f
or
N
L
P
ta
s
k
s
in
th
e
s
c
ie
nt
if
ic
dom
a
in
,
S
c
iB
E
R
T
[
8]
,
a
pr
e
tr
a
in
e
d
la
ngua
ge
m
ode
l
bui
lt
upon
B
E
R
T
,
i
s
in
tr
oduc
e
d.
L
e
ve
r
a
gi
ng
uns
up
e
r
vi
s
e
d
pr
e
tr
a
in
in
g
on
a
di
ve
r
s
e
,
m
ul
ti
di
s
c
ip
li
na
r
y
c
ol
le
c
ti
on
of
s
c
ie
nt
if
ic
publ
ic
a
ti
ons
a
nd
c
it
a
ti
ons
e
nha
nc
e
s
it
s
e
f
f
e
c
ti
ve
ne
s
s
on
dow
ns
tr
e
a
m
s
c
ie
nt
if
ic
N
L
P
ta
s
ks
.
E
m
pl
oyi
ng
i
ts
ow
n
voc
a
bul
a
r
y
c
a
ll
e
d
S
c
iVoc
a
b,
c
ons
is
ti
ng
of
s
c
ie
nt
if
ic
te
r
m
s
,
S
c
iB
E
R
T
s
how
s
s
ig
ni
f
ic
a
nt
s
ta
ti
s
ti
c
a
l
e
nha
nc
e
m
e
nt
s
c
om
pa
r
e
d
to
B
E
R
T
on
di
ve
r
s
e
s
c
ie
nt
if
ic
N
L
P
a
s
s
ig
nm
e
nt
s
,
e
nc
om
pa
s
s
in
g
s
e
qu
e
nc
e
ta
ggi
ng,
s
e
nt
e
nc
e
c
a
te
gor
iz
a
ti
on,
a
nd
de
pe
nde
nc
y
a
na
ly
s
is
.
W
hi
le
S
c
iB
E
R
T
e
f
f
e
c
ti
ve
ly
ha
ndl
e
s
ge
ne
r
a
l
s
c
ie
nt
if
ic
l
it
e
r
a
tu
r
e
, i
ts
pe
r
f
or
m
a
nc
e
i
n t
he
b
io
m
e
di
c
a
l
dom
a
in
i
s
l
im
i
te
d
due
t
o voc
a
bul
a
r
y c
ons
tr
a
in
ts
a
nd l
a
c
k of
s
pe
c
if
ic
bi
om
e
di
c
a
l
tr
a
in
in
g da
ta
. M
e
di
c
a
lB
E
R
T
, on the
ot
he
r
ha
nd,
is
pr
e
tr
a
in
e
d
s
pe
c
if
ic
a
ll
y
on
bi
om
e
di
c
a
l
c
or
pus
.
T
hi
s
s
pe
c
ia
li
z
a
ti
on
a
ll
ow
s
M
e
di
c
a
lB
E
R
T
to
a
c
hi
e
ve
hi
ghe
r
a
c
c
ur
a
c
y
in
ta
s
ks
s
uc
h
a
s
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
a
nd
r
e
la
ti
on
e
xt
r
a
c
ti
on,
pa
r
ti
c
ul
a
r
ly
on
da
ta
s
e
ts
li
ke
B
C
5C
D
R
-
C
h
e
m
ic
a
l
a
nd
N
C
B
I
-
D
is
e
a
s
e
,
w
he
r
e
it
out
pe
r
f
or
m
s
S
c
iB
E
R
T
by
a
n
F
1
m
a
r
gi
n
of
1.37
on
a
ve
r
a
ge
.
M
e
di
c
a
lB
E
R
T
’
s
dom
a
in
-
s
pe
c
if
ic
voc
a
bul
a
r
y
e
nha
nc
e
s
it
s
unde
r
s
ta
ndi
ng
of
c
om
pl
e
x
bi
om
e
di
c
a
l
te
r
m
s
,
a
ddr
e
s
s
in
g a
not
a
bl
e
g
a
p i
n S
c
iB
E
R
T
’
s
d
e
s
ig
n f
or
s
uc
h a
ppl
ic
a
t
io
ns
.
2.5. RoB
E
R
T
a
R
obus
tl
y
opt
im
iz
e
d
B
E
R
T
(
R
oB
E
R
T
a
)
[
9]
is
a
B
E
R
T
m
ode
l
th
a
t
unde
r
w
e
nt
tr
a
in
in
g
on
a
n
e
xpa
nde
d
E
ngl
is
h
da
ta
s
e
t
a
nd
f
or
a
lo
nge
r
pe
r
io
d
of
ti
m
e
us
in
g
s
e
lf
-
s
upe
r
vi
s
e
d
tr
a
in
in
g
te
c
hni
qu
e
s
.
T
hi
s
r
e
s
ul
ts
in
a
b
e
tt
e
r
m
ode
l
f
or
va
r
io
us
N
L
P
ta
s
k
s
.
R
o
B
E
R
T
a
’
s
e
xt
e
ns
iv
e
pr
e
tr
a
in
in
g
on
ge
n
e
r
a
l
E
ngl
is
h
te
xt
a
ll
ow
s
it
to
pe
r
f
or
m
r
e
a
s
ona
bl
y
w
e
ll
on
br
oa
d
bi
om
e
di
c
a
l
ta
s
ks
;
how
e
ve
r
,
it
f
a
ll
s
s
hor
t
in
de
e
pe
r
dom
a
in
-
s
pe
c
if
ic
c
ont
e
xt
s
.
F
or
in
s
ta
nc
e
,
w
hi
le
R
oB
E
R
T
a
de
m
on
s
tr
a
te
s
c
om
pe
ti
ti
ve
pe
r
f
or
m
a
nc
e
on
th
e
L
I
N
N
A
E
U
S
da
ta
s
e
t
w
it
h
a
n
F
1
s
c
or
e
of
87.8,
it
unde
r
pe
r
f
or
m
s
on
m
or
e
c
om
pl
e
x
bi
om
e
di
c
a
l
ta
s
k
s
.
M
e
di
c
a
lB
E
R
T
,
in
c
ont
r
a
s
t,
is
f
in
e
-
tu
ne
d
s
pe
c
if
ic
a
ll
y
f
or
bi
om
e
di
c
a
l
te
xt
m
in
in
g
,
out
pe
r
f
or
m
in
g
R
oB
E
R
T
a
in
ta
s
ks
r
e
qui
r
in
g
hi
ghe
r
c
ont
e
xt
ua
l
a
nd
dom
a
in
-
s
pe
c
if
ic
c
om
pr
e
he
ns
io
n,
s
uc
h
a
s
in
th
e
B
C
5
C
D
R
-
D
is
e
a
s
e
da
t
a
s
e
t,
w
h
e
r
e
M
e
di
c
a
lB
E
R
T
s
how
s
a
2.01
F
1
s
c
or
e
im
pr
ove
m
e
nt
ove
r
R
oB
E
R
T
a
.
T
hi
s
im
pr
ove
m
e
nt
unde
r
s
c
or
e
s
M
e
di
c
a
lB
E
R
T
’
s
a
dva
nt
a
g
e
i
n l
e
ve
r
a
gi
ng biom
e
di
c
a
l
te
r
m
in
ol
o
gy a
nd t
r
a
ns
f
e
r
l
e
a
r
ni
ng f
or
s
pe
c
ia
li
z
e
d t
a
s
k
s
.
2.6. Ot
h
e
r
w
or
k
s
B
r
e
s
s
e
m
e
t
al
.
[
10]
in
tr
oduc
e
d,
a
pr
e
-
tr
a
in
e
d
B
E
R
T
m
ode
l
us
in
g
a
la
r
ge
c
or
pus
of
G
e
r
m
a
n
m
e
di
c
a
l
doc
um
e
nt
s
,
in
c
lu
di
ng
r
a
di
ol
ogy
r
e
por
ts
,
P
ubM
e
d
a
bs
tr
a
c
ts
,
S
pr
in
ge
r
N
a
tu
r
e
,
a
nd
G
e
r
m
a
n
W
ik
ip
e
di
a
.
P
r
e
pr
oc
e
s
s
in
g
s
t
e
ps
in
vol
ve
d
da
ta
a
nonymi
z
a
ti
on
(
r
e
m
ovi
ng
pa
ti
e
nt
na
m
e
s
w
it
h
a
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
m
ode
l)
a
nd
de
dupl
ic
a
ti
on
(
us
in
g
c
o
s
in
e
s
im
il
a
r
it
y)
.
T
he
m
ode
l
a
c
hi
e
ve
d s
ta
te
-
of
-
th
e
-
a
r
t
pe
r
f
or
m
a
nc
e
on e
ig
ht
m
e
di
c
a
l
be
nc
hm
a
r
ks
,
la
r
ge
ly
due
to
th
e
e
xt
e
ns
iv
e
tr
a
in
in
g
da
t
a
,
w
it
h
e
f
f
ic
ie
nt
to
ke
ni
z
a
ti
on
ha
vi
ng
a
le
s
s
e
r
e
f
f
e
c
t
on r
e
s
ul
ts
.
W
a
da
e
t
al
.
[
11]
ta
c
kl
e
d
B
E
R
T
'
s
w
e
a
k
e
r
pe
r
f
or
m
a
nc
e
on
s
m
a
ll
e
r
bi
om
e
di
c
a
l
c
or
por
a
by
up
-
s
a
m
pl
in
g
s
m
a
ll
e
r
da
ta
s
e
ts
a
nd
pr
e
-
tr
a
in
in
g
on
bot
h
la
r
ge
a
nd
s
m
a
ll
c
or
por
a
.
T
hi
s
in
vol
ve
d
s
e
gm
e
nt
in
g
bot
h
c
or
por
a
in
to
s
im
il
a
r
-
s
iz
e
d
doc
um
e
nt
s
a
nd
in
c
r
e
a
s
in
g
th
e
s
m
a
ll
e
r
c
or
pus
to
m
a
tc
h
th
e
la
r
ge
r
one
.
T
hr
e
e
e
xpe
r
im
e
nt
s
w
e
r
e
c
onduc
te
d:
a
J
a
pa
ne
s
e
m
e
di
c
a
l
B
E
R
T
,
a
n
E
ngl
is
h
bi
om
e
di
c
a
l
B
E
R
T
,
a
nd
a
n
e
nha
nc
e
d
bi
om
e
di
c
a
l
B
E
R
T
f
r
om
P
ubM
e
d.
R
e
s
ul
ts
s
how
e
d
th
e
J
a
pa
n
e
s
e
B
E
R
T
e
x
c
e
ll
e
d
in
m
e
di
c
a
l
c
la
s
s
if
ic
a
ti
on
ta
s
ks
,
w
hi
le
th
e
E
ngl
is
h
B
E
R
T
pe
r
f
or
m
e
d
w
e
ll
on
th
e
B
L
U
E
be
nc
hm
a
r
k,
w
it
h
th
e
e
nha
nc
e
d
B
E
R
T
m
ode
l
im
pr
ovi
ng s
c
or
e
s
by 0.3 point
s
ove
r
t
he
a
bl
a
ti
on mode
l.
3.
O
V
E
R
V
I
E
W OF
B
E
R
T
A
N
D
B
E
N
C
H
M
A
R
K
S
3.1. B
E
R
T
B
E
R
T
a
r
e
a
tr
a
ns
f
or
m
e
r
-
de
r
iv
e
d
la
ngua
ge
r
e
pr
e
s
e
nt
a
ti
on
m
ode
l
de
s
ig
ne
d
to
unde
r
s
ta
nd
a
nd
c
a
pt
ur
e
th
e
c
ont
e
xt
ua
l
s
ig
ni
f
ic
a
nc
e
of
w
or
ds
w
it
hi
n
a
pr
ovi
de
d
s
e
n
te
nc
e
or
pa
s
s
a
ge
by
c
a
pt
ur
in
g
bi
di
r
e
c
ti
ona
l
r
e
la
ti
ons
hi
ps
be
twe
e
n
w
or
ds
[
1]
.
B
E
R
T
e
m
pl
oys
a
tr
a
ns
f
or
m
e
r
a
r
c
hi
te
c
tu
r
e
ba
s
e
d
on
s
e
lf
-
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
[
12]
.
C
ont
r
a
r
y
to
r
e
c
ur
r
e
nt
n
e
ur
a
l
ne
twor
ks
(
R
N
N
s
)
,
lo
ng
s
hor
t
-
te
r
m
m
e
m
or
y
(
L
S
T
M
)
,
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2367
-
2378
2370
ot
he
r
m
ode
ls
ba
s
e
d
on
uni
di
r
e
c
ti
ona
l
a
tt
e
nt
io
n
th
a
t
s
c
r
ut
in
iz
e
te
xt
e
it
he
r
f
r
om
le
f
t
to
r
ig
ht
or
f
r
om
r
ig
ht
to
le
f
t
[
13]
,
[
14]
,
th
e
a
r
c
hi
te
c
tu
r
e
of
B
E
R
T
e
na
bl
e
s
it
to
und
e
r
s
ta
n
d
th
e
r
e
la
ti
ons
hi
p
s
be
tw
e
e
n
w
or
d
s
w
it
hi
n
a
s
e
nt
e
nc
e
by pr
oc
e
s
s
in
g t
e
xt
i
n both di
r
e
c
ti
ons
, bot
h l
e
f
twa
r
d a
n
d r
ig
ht
w
a
r
d, by p
r
oc
e
s
s
in
g t
he
e
nt
ir
e
s
e
que
nc
e
s
im
ul
ta
ne
ous
ly
.
B
y
e
m
pl
oyi
ng
two
uns
up
e
r
vi
s
e
d
le
a
r
ni
ng
ta
s
k
s
,
na
m
e
ly
m
a
s
ke
d
la
ngua
g
e
m
ode
li
ng
(
M
L
M
)
a
nd
ne
xt
s
e
nt
e
nc
e
pr
e
di
c
ti
on
(
N
S
P
)
,
B
E
R
T
c
a
n
e
nga
ge
in
pr
e
tr
a
in
in
g
on
e
xt
e
ns
iv
e
a
m
ount
s
of
unl
a
be
le
d
te
xt
ta
il
or
e
d t
o a
s
pe
c
if
ic
doma
in
[
1]
.
I
n
M
L
M
,
e
a
c
h
to
ke
n
i
s
c
onv
e
r
te
d
in
to
w
or
d
e
m
be
ddi
ng
s
.
I
n
th
e
in
put
s
e
que
n
c
e
,
a
c
e
r
ta
in
p
e
r
c
e
nt
a
ge
(
~
15%
)
of
th
e
in
put
to
ke
n
s
a
r
e
r
a
ndoml
y
c
ho
s
e
n a
nd
obs
c
ur
e
d
t
o
be
s
ubs
ti
tu
te
d w
it
h
a
s
pe
c
i
a
l
[
M
A
S
K
]
to
ke
n
(
~
80%
)
.
T
he
m
ode
l
ha
ndl
e
s
th
e
in
put
s
e
qu
e
nc
e
to
ke
n
by
to
ke
n
us
in
g
s
e
lf
-
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
a
nd
m
ul
ti
la
ye
r
ne
ur
a
l
ne
twor
ks
[
13]
.
F
or
e
a
c
h m
a
s
ke
d t
oke
n, t
he
m
ode
l
ge
ne
r
a
te
s
a
pr
oba
bi
li
ty
di
s
tr
ib
ut
io
n a
c
r
os
s
th
e
e
nt
ir
e
to
ke
n
voc
a
bul
a
r
y.
B
y
tr
a
in
in
g
th
e
m
ode
l
to
r
e
c
on
s
tr
uc
t
th
e
m
a
s
ke
d
to
ke
ns
,
th
e
m
ode
l
le
a
r
n
s
to
pr
oduc
e
c
ohe
r
e
nt
r
e
pr
e
s
e
nt
a
ti
ons
of
t
he
t
e
xt
[
1]
. N
S
P
de
te
r
m
in
e
s
w
he
th
e
r
t
w
o s
e
nt
e
nc
e
s
a
ppe
a
r
c
ons
e
c
ut
iv
e
ly
in
a
te
xt
c
or
pus
.
T
he
B
E
R
T
m
ode
l
ta
gs
e
a
c
h
in
s
t
a
nc
e
to
de
n
ot
e
w
he
th
e
r
th
e
s
ubs
e
que
nt
s
e
nt
e
n
c
e
lo
gi
c
a
ll
y
s
uc
c
e
e
d
s
th
e
pr
e
c
e
di
ng
s
e
nt
e
nc
e
in
th
e
s
our
c
e
te
xt
.
I
t
to
ke
ni
z
e
s
th
e
s
e
nt
e
nc
e
s
a
nd
c
r
e
a
te
s
e
m
be
ddi
ng
s
f
or
e
a
c
h
to
ke
n.
S
pe
c
i
a
l
to
ke
ns
s
uc
h
a
s
[
C
L
S
]
(
c
la
s
s
if
ic
a
ti
on)
a
nd
[
S
E
P
]
(
s
e
pa
r
a
to
r
)
a
r
e
in
s
e
r
te
d
be
twe
e
n
s
e
nt
e
nc
e
s
, a
nd s
e
gm
e
nt
e
m
be
ddi
ngs
a
r
e
i
nt
r
oduc
e
d t
o di
f
f
e
r
e
nt
ia
te
be
twe
e
n t
he
t
w
o
s
e
nt
e
nc
e
s
[
1]
.
3.2. Dow
n
s
t
r
e
am
n
at
u
r
al
l
an
gu
age
p
r
oc
e
s
s
in
g
t
as
k
s
3.2.1. Nam
e
d
e
n
t
it
y
r
e
c
ogn
it
io
n
C
onduc
ti
ng
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
on
bi
om
e
di
c
a
l
da
ta
s
e
t
s
in
vol
ve
s
id
e
nt
if
yi
ng
a
nd
c
la
s
s
if
yi
ng
e
nt
it
ie
s
li
ke
ge
ne
s
,
di
s
e
a
s
e
s
,
pr
ot
e
in
s
,
a
nd
c
h
e
m
ic
a
ls
pr
e
s
e
nt
in
t
he
te
xt
.
O
ld
e
r
m
ode
ls
in
th
is
dom
a
in
, s
uc
h
a
s
c
ondi
ti
ona
l
r
a
ndom
f
ie
ld
s
(
C
R
F
s
)
[
15
]
,
[
16
]
a
nd
bi
di
r
e
c
ti
ona
l
l
ong
s
hor
t
-
te
r
m
m
e
m
o
r
y
(
B
i
-
L
S
T
M
)
[
17]
,
[
18
]
ne
twor
ks
,
de
pe
nd
on
m
a
nua
ll
y
c
r
a
f
te
d
f
e
a
tu
r
e
s
or
s
e
que
nt
ia
l
de
pe
nde
nc
ie
s
but
la
c
k
th
e
a
bi
li
ty
to
c
a
pt
ur
e
in
tr
ic
a
te
c
ont
e
xt
r
e
pr
e
s
e
nt
a
ti
ons
e
f
f
e
c
ti
ve
ly
.
B
E
R
T
out
pe
r
f
or
m
s
th
e
s
e
ol
de
r
m
ode
ls
by
le
ve
r
a
gi
ng
bi
di
r
e
c
ti
ona
l
c
ont
e
xt
unde
r
s
ta
ndi
ng
a
nd
tr
a
ns
f
or
m
e
r
a
r
c
hi
te
c
tu
r
e
s
.
T
he
m
ode
l
le
a
r
n
s
to
la
be
l
e
a
c
h
in
pu
t
to
ke
n
w
it
h
it
s
c
or
r
e
s
ponding
e
nt
it
y
la
be
l
or
e
nt
it
y t
ype
us
in
g B
I
O
t
a
gs
(
be
gi
nni
ng, i
ns
id
e
, out
s
id
e
)
, w
hi
c
h
o
ur
m
ode
l
e
m
pl
oys
f
or
t
h
is
t
a
s
k. T
he
F
1 m
e
tr
ic
a
nd w
or
d
-
le
ve
l
m
a
c
r
o F
1 (
a
ppl
ie
d t
o P
I
C
O
t
a
s
k
[
19]
)
w
e
r
e
ut
il
i
z
e
d f
or
r
e
s
ul
t
e
va
lu
a
ti
on.
3.2.2. Re
la
t
io
n
e
xt
r
ac
t
io
n
I
n
th
e
bi
om
e
di
c
a
l
dom
a
in
,
r
e
la
ti
on
e
xt
r
a
c
ti
on
in
vol
ve
s
id
e
nt
if
yi
ng
a
nd
c
a
te
gor
iz
in
g
c
onne
c
ti
ons
a
m
ong
e
nt
it
ie
s
m
e
nt
io
ne
d
in
te
xt
ua
l
c
ont
e
nt
,
s
u
c
h
a
s
pr
ot
e
i
ns
,
ge
ne
s
,
di
s
e
a
s
e
s
,
a
nd
dr
ug
s
.
M
e
th
od
s
li
ke
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
s
(
S
V
M
s
)
a
nd C
R
F
s
[
15]
,
[
16]
,
or
c
onv
ol
ut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
s
)
e
nde
a
vor
to
pe
r
f
or
m
r
e
la
ti
on
e
xt
r
a
c
ti
on
by
le
ve
r
a
gi
ng
li
ngui
s
ti
c
f
e
a
tu
r
e
s
,
s
ynt
a
c
ti
c
p
a
r
s
in
g,
or
s
e
que
n
c
e
m
ode
li
ng.
H
ow
e
ve
r
, t
he
y s
tr
uggl
e
d t
o c
a
pt
ur
e
c
om
pl
e
x
s
e
m
a
nt
ic
r
e
la
ti
ons
hi
ps
e
f
f
e
c
ti
ve
ly
.
A
nnot
a
te
d
da
ta
s
e
ts
w
it
h
e
nt
it
ie
s
@
G
E
N
E
$
a
nd
@
D
I
S
E
A
S
E
$
in
th
e
ge
ne
ti
c
a
s
s
oc
ia
ti
on
da
ta
ba
s
e
(
G
A
D
)
a
nd
E
ur
ope
a
n U
ni
on a
dve
r
s
e
dr
ug r
e
a
c
ti
ons
(
EU
-
ADR
)
da
ta
s
e
ts
, @
G
E
N
E
a
nd @
C
H
E
M
I
C
A
L
$ i
n t
he
C
he
m
P
r
ot
da
ta
s
e
t
a
nd
@
D
R
U
G
$
in
th
e
D
D
I
da
ta
s
e
t
a
r
e
e
m
pl
oye
d
f
or
tr
a
in
in
g
pur
pos
e
s
.
E
a
c
h
s
e
que
nc
e
is
f
or
m
a
tt
e
d
w
it
h
s
pe
c
ia
l
to
ke
n
s
,
a
ll
ow
in
g
B
E
R
T
to
unde
r
s
t
a
nd
th
e
r
e
la
ti
ons
hi
ps
b
e
twe
e
n
e
nt
it
ie
s
.
T
h
e
s
e
uni
que
to
ke
ns
de
not
e
th
e
s
ta
r
t
a
nd
e
nd
o
f
e
nt
it
ie
s
or
r
e
la
ti
ons
hi
ps
,
na
m
e
ly
[
C
L
S
]
a
nd
[
S
E
P
]
.
T
he
F
1
a
nd
m
ic
r
o
a
ve
r
a
ge
-
F
1 m
e
tr
ic
s
a
r
e
e
m
pl
oye
d t
o a
s
s
e
s
s
t
he
out
c
om
e
s
of
t
hi
s
t
a
s
k.
3.
2
.3. Qu
e
s
t
io
n
a
n
s
w
e
r
in
g
Q
ue
s
ti
on
a
n
s
w
e
r
in
g
w
it
hi
n
th
e
bi
om
e
di
c
a
l
dom
a
in
e
nc
om
p
a
s
s
e
s
pr
ovi
di
ng
a
n
s
w
e
r
s
to
que
r
ie
s
f
or
m
ul
a
te
d
in
na
tu
r
a
l
la
ngua
ge
,
r
e
ly
in
g
on
in
f
or
m
a
ti
on
c
ont
a
in
e
d
w
it
hi
n
bi
om
e
di
c
a
l
te
xt
s
li
ke
r
e
s
e
a
r
c
h
a
r
ti
c
le
s
a
nd
c
li
ni
c
a
l
r
e
c
or
ds
.
I
nf
or
m
a
ti
on
r
e
tr
ie
va
l
s
ys
te
m
s
a
n
d
th
e
J
a
va
li
br
a
r
y
L
uc
e
ne
ut
il
iz
e
d
te
c
hni
que
s
s
uc
h
a
s
te
r
m
f
r
e
que
nc
y
-
in
ve
r
s
e
doc
um
e
nt
f
r
e
que
nc
y
(
TF
-
I
D
F
)
a
nd
s
tr
a
ig
ht
f
or
w
a
r
d
to
ke
n
m
a
tc
hi
ng
f
o
r
c
onduc
ti
ng que
s
ti
on a
ns
w
e
r
in
g, but t
he
y w
e
r
e
l
im
it
e
d i
n c
a
pt
ur
in
g c
ont
e
xt
ua
l
a
nd s
e
m
a
nt
ic
und
e
r
s
ta
ndi
ng.
B
E
R
T
le
a
r
n
s
by
pr
oc
e
s
s
in
g
bot
h
th
e
que
s
ti
on
s
a
nd
c
ont
e
xt
pa
s
s
a
ge
s
th
r
ough
it
s
tr
a
n
s
f
or
m
e
r
-
ba
s
e
d
a
r
c
hi
te
c
tu
r
e
,
in
vol
vi
ng
s
e
ve
r
a
l
la
ye
r
s
of
s
e
lf
-
a
tt
e
nt
io
n
a
nd
e
nc
o
di
ng
m
e
c
ha
ni
s
m
s
.
W
e
us
e
d
two
da
ta
s
e
ts
f
r
om
bi
om
e
di
c
a
l
la
ngua
ge
unde
r
s
ta
ndi
ng
a
nd
r
e
a
s
oni
ng
be
nc
hm
a
r
k
(
B
L
U
R
B
)
[
20
]
:
P
ubM
e
d
que
s
ti
on
a
ns
w
e
r
in
g
(
P
ubM
e
dQ
A
)
[
21]
a
nd
bi
om
e
di
c
a
l
s
e
m
a
nt
ic
in
de
xi
ng
a
nd
que
s
ti
on
-
a
ns
w
e
r
in
g
c
ha
ll
e
nge
(
B
io
A
S
Q
)
[
22]
to
pe
r
f
or
m
t
he
que
s
ti
on
-
a
ns
w
e
r
in
g t
a
s
k a
nd e
va
lu
a
t
e
d t
he
r
e
s
ul
ts
b
a
s
e
d on the
ir
a
c
c
ur
a
c
y.
3.
2
.4. S
e
n
t
e
n
c
e
si
m
il
ar
it
y
S
e
nt
e
nc
e
s
im
il
a
r
it
y
in
vol
ve
s
e
n
c
odi
ng
e
a
c
h
s
e
nt
e
n
c
e
s
e
pa
r
a
te
ly
w
it
h
B
E
R
T
,
e
xt
r
a
c
ti
ng
th
e
e
m
be
ddi
ngs
c
or
r
e
s
ponding
to
e
a
c
h
to
ke
n,
a
nd
a
ggr
e
ga
ti
ng
th
e
s
e
e
m
be
ddi
ngs
in
to
a
f
ix
e
d
-
s
iz
e
ve
c
to
r
r
e
pr
e
s
e
nt
a
ti
on
e
nc
om
pa
s
s
in
g
th
e
e
nt
ir
e
s
e
nt
e
n
c
e
.
A
f
te
r
th
e
s
e
nt
e
nc
e
e
m
b
e
ddi
ngs
a
r
e
a
c
qui
r
e
d,
c
o
s
in
e
s
im
il
a
r
it
y
is
e
m
pl
oye
d
to
ga
uge
th
e
s
im
il
a
r
it
y
be
twe
e
n
th
e
tw
o
ve
c
to
r
s
.
A
hi
ghe
r
s
c
or
e
f
or
c
os
in
e
s
im
il
a
r
it
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
e
di
c
al
B
E
R
T
:
e
nhanc
in
g bi
om
e
di
c
al
nat
ur
al
l
anguage
pr
oc
e
s
s
in
g us
in
g pr
e
tr
ai
ne
d
…
(
K
. Sahit
R
e
ddy
)
2371
in
di
c
a
te
s
gr
e
a
te
r
s
im
il
a
r
it
y
be
twe
e
n
s
e
nt
e
nc
e
s
,
w
hi
le
a
lo
w
e
r
s
c
or
e
im
pl
ie
s
le
s
s
s
im
il
a
r
it
y.
W
e
e
m
pl
oye
d
th
e
bi
om
e
di
c
a
l
s
e
m
a
nt
ic
s
e
nt
e
nc
e
s
im
il
a
r
it
y
e
s
ti
m
a
ti
on
s
ys
te
m
(
B
I
O
S
S
E
S
)
[
23
]
da
ta
s
e
t
f
r
om
B
L
U
R
B
a
nd
e
va
lu
a
te
d t
he
pe
r
f
or
m
a
nc
e
on t
he
P
e
a
r
s
on me
tr
ic
.
3.
2
.5. D
oc
u
m
e
n
t
m
u
lt
il
ab
e
l
c
la
s
s
i
f
ic
at
io
n
D
oc
um
e
nt
c
la
s
s
if
ic
a
ti
on
i
s
a
n
N
L
P
ta
s
k
th
a
t
in
vol
ve
s
a
s
s
ig
ni
ng
c
a
te
gor
ie
s
or
c
la
s
s
e
s
to
a
do
c
um
e
nt
.
T
hi
s
m
e
th
od
s
im
pl
if
ie
s
th
e
m
a
na
g
e
m
e
nt
,
s
e
a
r
c
hi
ng,
f
il
te
r
in
g,
a
nd
a
na
ly
s
is
of
doc
um
e
nt
s
.
W
e
e
v
a
lu
a
te
d
th
e
be
nc
hm
a
r
k on the
ha
ll
m
a
r
ks
of
c
a
nc
e
r
(
H
oC
)
da
ta
s
e
t
[
24]
, w
hi
c
h i
s
a
ls
o a
va
il
a
bl
e
i
n B
L
U
R
B
[
20]
.
3.3. Dat
as
e
t
s
an
d
b
e
n
c
h
m
ar
k
s
W
e
ut
il
iz
e
d e
xi
s
ti
ng da
ta
s
e
ts
us
e
d f
or
t
a
s
ks
i
n t
he
B
L
U
R
B
[
20]
be
nc
hm
a
r
k a
nd B
io
B
E
R
T
[
4]
, w
hi
c
h
a
r
e
c
om
m
onl
y
us
e
d
by
th
e
B
io
N
L
P
c
om
m
uni
ty
.
A
to
ta
l
of
17
da
ta
s
e
ts
w
e
r
e
us
e
d,
a
nd
T
a
bl
e
1
li
s
t
s
th
e
tr
a
in
in
g
a
nd
te
s
ti
ng
in
s
ta
nc
e
s
f
or
e
a
c
h
of
th
e
f
iv
e
ta
s
ks
,
w
hi
c
h
in
c
lu
de
s
e
qu
e
nc
e
la
b
e
li
ng,
s
e
que
n
c
e
c
la
s
s
if
ic
a
ti
on, que
s
ti
on a
ns
w
e
r
in
g, s
e
m
a
nt
ic
s
e
nt
e
nc
e
s
im
il
a
r
it
y, a
nd doc
um
e
nt
c
la
s
s
if
ic
a
ti
on.
T
a
bl
e
1. S
ta
ti
s
ti
c
s
of
t
he
d
a
ta
s
e
t
s
ut
il
iz
e
d f
or
f
in
e
-
tu
ni
ng. T
he
t
r
a
in
in
g a
nd t
e
s
t
c
ol
um
ns
i
ndi
c
a
t
e
t
he
numbe
r
of
t
r
a
in
in
g a
nd t
e
s
ti
ng e
xa
m
pl
e
s
f
or
e
a
c
h da
ta
s
e
t
r
e
s
pe
c
ti
ve
ly
D
a
t
a
s
e
t
T
a
s
k
T
r
a
i
n
T
e
s
t
N
C
B
I
-
di
s
e
a
s
e
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
5134
960
B
C
2G
M
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
15197
6325
B
C
5C
D
R
-
di
s
e
a
s
e
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
4182
4424
B
C
5C
D
R
-
c
h
e
m
i
c
a
l
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
5203
5385
J
N
L
P
B
A
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
46750
8662
L
I
N
N
A
E
U
S
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
on
281273
165095
B
C
4C
H
E
M
D
N
a
m
e
d
e
nt
i
t
y r
e
c
ogni
t
i
o
n
893685
767636
S
pe
c
i
e
s
-
800
N
a
m
e
d
e
nt
i
t
y r
e
c
og
ni
t
i
on
147291
42298
E
B
M
-
P
I
C
O
P
I
C
O
339167
16364
GAD
R
e
l
a
t
i
on
e
xt
r
a
c
t
i
on
4261
535
EU
-
ADR
R
e
l
a
t
i
on
e
xt
r
a
c
t
i
on
3195
355
DDI
-
2013
R
e
l
a
t
i
on
e
xt
r
a
c
t
i
on
22233
5716
C
he
m
P
r
ot
R
e
l
a
t
i
on
e
xt
r
a
c
t
i
on
18035
15745
P
ubM
e
dQ
A
Q
ue
s
t
i
on
a
ns
w
e
r
i
ng
450
500
B
i
oA
S
Q
Q
ue
s
t
i
on
a
ns
w
e
r
i
ng
670
140
B
I
O
S
S
E
S
S
e
nt
e
nc
e
s
i
m
i
l
a
r
i
t
y
64
20
H
oC
M
ul
t
i
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on
1295
371
N
C
B
I
-
di
s
e
a
s
e
[
25]
is
e
nt
ir
e
ly
a
nnot
a
te
d
a
nd
c
ont
a
in
s
793
a
bs
tr
a
c
ts
f
r
om
P
ubM
e
d,
f
e
a
tu
r
in
g
6,892
m
e
nt
io
ns
of
di
s
e
a
s
e
s
a
nd
790
di
s
ti
nc
t
di
s
e
a
s
e
c
onc
e
pt
s
.
T
h
e
B
io
C
r
e
a
ti
ve
V
C
D
R
(
B
C
5C
D
R
)
[
26]
da
ta
s
e
t
c
om
pr
is
e
s
le
ngt
hy
doc
um
e
nt
s
th
a
t
a
r
e
s
e
gm
e
nt
e
d
in
to
s
e
nt
e
n
c
e
s
to
m
it
ig
a
te
th
e
ir
s
iz
e
,
a
s
th
e
y
c
a
nnot
be
di
r
e
c
tl
y
pr
oc
e
s
s
e
d
by
l
a
ngua
ge
m
od
e
ls
du
e
to
s
iz
e
c
ons
tr
a
in
ts
.
T
he
two
e
nt
it
y
ty
pe
s
a
r
e
c
he
m
ic
a
l
a
nd
di
s
e
a
s
e
.
T
he
B
io
C
r
e
a
ti
ve
I
V
c
h
e
m
ic
a
l
c
om
pound
a
nd
dr
ug
na
m
e
r
e
c
o
gni
ti
on
(
B
C
4C
H
E
M
D
)
[
19]
da
ta
ba
s
e
c
ont
a
in
s
10,000 P
ubM
e
d a
bs
tr
a
c
ts
t
ha
t
c
ont
a
in
84,355 c
he
m
ic
a
l
e
nt
it
y m
e
nt
io
ns
i
n t
ot
a
l.
T
he
y a
r
e
l
a
be
le
d m
a
nua
ll
y by
c
he
m
is
tr
y l
it
e
r
a
tu
r
e
c
ur
a
to
r
s
.
T
he
B
io
C
r
e
a
ti
ve
2
ge
ne
m
e
nt
io
n
(
B
C
2G
M
)
[
27
]
c
or
pus
is
a
n
a
ggr
e
ga
ti
on
of
va
r
io
us
s
e
nt
e
nc
e
s
,
w
he
r
e
e
a
c
h
s
e
nt
e
nc
e
is
c
om
pos
e
d
of
ge
ne
m
e
nt
io
ns
(
G
E
N
E
a
nnot
a
ti
ons
)
.
P
a
r
ti
c
ip
a
nt
s
a
r
e
ta
s
ke
d
w
it
h
pi
npoi
nt
in
g
a
ge
ne
m
e
nt
io
n
w
it
hi
n
a
s
e
nt
e
nc
e
by
id
e
nt
if
yi
ng
th
e
s
ta
r
ti
ng
a
nd
e
ndi
ng
c
ha
r
a
c
te
r
s
of
th
a
t
s
e
nt
e
nc
e
.
J
N
L
P
B
A
[
28]
is
de
r
iv
e
d
f
r
om
th
e
G
E
N
I
A
ve
r
s
io
n
3.
02
c
or
pus
.
T
hi
s
to
ol
w
a
s
c
ons
tr
uc
te
d
ba
s
e
d
on
a
c
ont
r
ol
le
d
s
e
a
r
c
h
of
M
E
D
L
I
N
E
us
in
g
M
e
S
H
te
r
m
s
s
uc
h
a
s
“
bl
ood
c
e
ll
s
”
a
nd
“
tr
a
ns
c
r
ip
ti
on
f
a
c
to
r
s
”
.
A
ppr
oxi
m
a
te
ly
2,000
a
bs
tr
a
c
ts
w
e
r
e
f
il
te
r
e
d
a
nd
a
nnot
a
te
d
by
ha
nd
in
a
c
c
or
da
nc
e
w
it
h
a
s
m
a
ll
ta
xonomy
of
48 c
la
s
s
e
s
ba
s
e
d on a
c
h
e
m
ic
a
l
c
la
s
s
if
ic
a
ti
on.
S
pe
c
ie
s
-
800
[
29]
is
a
c
or
pus
w
hos
e
a
b
s
tr
a
c
ts
a
r
e
m
a
nu
a
ll
y
a
nnot
a
te
d.
I
t
is
a
c
or
pus
f
or
s
pe
c
if
ic
e
nt
it
ie
s
.
I
t
c
ons
i
s
ts
of
800
P
ubM
e
d
a
bs
tr
a
c
ts
th
a
t
c
ont
a
in
or
ga
n
is
m
m
e
nt
io
ns
.
F
r
om
e
ig
ht
di
s
c
ip
li
ne
s
,
n
a
m
e
ly
ba
c
te
r
io
lo
gy,
bot
a
ny,
e
nt
om
ol
ogy,
m
e
di
c
in
e
,
m
yc
ol
ogy,
pr
ot
is
to
lo
gy,
vi
r
ol
ogy,
a
nd
z
ool
ogy,
100
a
b
s
tr
a
c
ts
w
e
r
e
c
hos
e
n.
L
I
N
N
A
E
U
S
[
30]
c
or
pus
in
c
lu
de
s
100
f
ul
l
-
te
xt
do
c
um
e
nt
s
s
our
c
e
d
r
a
ndoml
y
f
r
om
th
e
P
M
C
O
A
doc
um
e
nt
s
e
t.
S
pe
c
ie
s
w
it
hi
n
th
e
s
e
doc
um
e
nt
s
w
e
r
e
a
nnot
a
t
e
d
by
ha
nd
a
nd
th
e
n
s
ta
nda
r
di
z
e
d
to
th
e
N
C
B
I
ta
xonomy I
D
s
c
or
r
e
s
ponding t
o t
he
s
pe
c
if
ic
s
p
e
c
ie
s
m
e
nt
io
ne
d.
E
B
M
-
P
I
C
O
[
31]
us
e
d
f
or
th
e
P
I
C
O
ta
s
k
c
ont
a
in
s
4,993
a
bs
tr
a
c
ts
.
T
h
e
s
e
a
r
e
a
nnot
a
te
d
w
it
h
(
P
)
a
r
ti
c
ip
a
nt
s
,
(
I
)
n
te
r
ve
nt
io
n,
(
C
)
om
pa
r
a
to
r
,
a
nd
(
O
)
ut
c
om
e
s
.
G
A
D
[
32]
is
a
c
or
pus
th
a
t
is
us
e
d
to
id
e
nt
if
y
a
s
s
oc
ia
ti
on
s
be
twe
e
n
g
e
ne
s
a
nd
di
s
e
a
s
e
s
.
I
t
us
e
s
a
s
e
m
ia
ut
om
a
ti
c
a
nnot
a
ti
on
pr
oc
e
dur
e
.
E
U
-
A
D
R
[
33]
da
ta
s
e
t
ha
s
be
e
n
a
nnot
a
te
d
f
or
di
s
or
de
r
s
,
dr
ugs
,
th
e
ir
in
te
r
r
e
la
ti
ons
hi
ps
a
nd
ge
ne
s
.
T
o
und
e
r
s
ta
nd
th
e
s
e
r
e
la
ti
ons
hi
ps
in
te
xt
s
,
th
e
a
nnot
a
te
d
r
e
la
ti
ons
hi
ps
s
e
r
ve
a
s
a
ba
s
is
f
or
tr
a
in
in
g
a
nd
a
s
s
e
s
s
in
g
te
xt
m
in
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2367
-
2378
2372
te
c
hni
que
s
.
C
h
e
m
P
r
ot
[
34]
da
ta
s
e
t
id
e
nt
if
ie
s
c
he
m
ic
a
l
a
nd
pr
ot
e
in
e
nt
it
ie
s
a
nd
th
e
ir
li
ke
ly
r
e
la
ti
on
to
one
a
not
he
r
. T
he
s
e
c
om
pounds
a
r
e
g
e
ne
r
a
ll
y a
c
ti
va
to
r
s
or
i
nhi
bi
to
r
s
of
pr
ot
e
in
s
.
P
ubM
e
dQ
A
[
21]
is
a
da
ta
s
e
t
ba
s
e
d
on
bi
om
e
di
c
a
l
que
s
ti
on
a
ns
w
e
r
in
g
c
ol
le
c
te
d
f
r
om
P
ubM
e
d
a
bs
tr
a
c
ts
.
T
h
e
obj
e
c
ti
ve
of
th
e
m
ode
l
de
ve
lo
pe
d
w
it
h
th
is
da
ta
s
e
t
pr
ovi
de
s
a
ns
w
e
r
s
to
bi
om
e
di
c
a
l
que
s
ti
on
s
s
uc
h
a
s
Y
E
S
,
N
O
,
or
M
A
Y
B
E
,
ba
s
e
d
on
in
f
or
m
a
ti
on
f
r
o
m
th
e
a
bs
tr
a
c
ts
.
T
he
da
ta
s
e
t
ha
s
1000
e
xpe
r
t
-
a
nnot
a
te
d,
61,200
unl
a
be
le
d
a
nd
211,300
a
r
ti
f
ic
ia
ll
y
ge
ne
r
a
te
d
que
s
ti
on
a
ns
w
e
r
in
s
ta
nc
e
s
.
E
a
c
h
in
s
ta
nc
e
c
ons
is
ts
of
a
que
s
ti
on t
ha
t
is
ba
s
e
d on the
e
xi
s
ti
ng r
e
s
e
a
r
c
h t
it
le
.
B
io
A
S
Q
[
22]
que
s
ti
on a
ns
w
e
r
in
g da
ta
s
e
t
c
ons
is
ts
of
, i
n a
ddi
ti
o
n t
o e
xa
c
t
a
ns
w
e
r
s
, i
de
a
l
a
ns
w
e
r
s
t
ha
t
a
r
e
us
e
f
ul
f
or
r
e
s
e
a
r
c
h
on
m
ul
ti
doc
um
e
nt
s
um
m
a
r
iz
a
ti
on,
unl
ik
e
th
e
m
a
jo
r
it
y
of
th
e
pr
e
vi
ous
que
s
ti
on
a
ns
w
e
r
in
g
be
nc
hm
a
r
ks
, w
hi
c
h c
ons
i
s
t
of
onl
y e
xa
c
t
a
ns
w
e
r
s
. I
t
c
ons
is
ts
of
s
tr
uc
tu
r
e
d a
nd un
s
tr
uc
tu
r
e
d da
ta
. I
t
c
ons
is
ts
of
doc
um
e
nt
s
a
nd
s
ni
pp
e
ts
th
a
t
pr
ove
us
e
f
ul
f
or
r
e
tr
ie
va
l
of
in
f
o
r
m
a
ti
on
a
nd
pa
s
s
a
ge
r
e
tr
ie
va
l
e
xpe
r
im
e
nt
s
.
B
I
O
S
S
E
S
[
23]
is
a
be
nc
hm
a
r
k
da
ta
s
e
t
f
or
bi
om
e
di
c
a
l
s
e
nt
e
n
c
e
s
im
il
a
r
it
y
e
s
ti
m
a
ti
on.
T
he
da
ta
s
e
t
c
ont
a
in
s
100
s
e
nt
e
nc
e
pa
ir
s
,
s
e
le
c
te
d
f
r
om
th
e
te
xt
a
na
ly
s
is
c
onf
e
r
e
nc
e
(
T
A
C
)
bi
om
e
di
c
a
l
s
um
m
a
r
iz
a
ti
on
tr
a
c
k
tr
a
in
in
g
da
ta
s
e
t
,
w
hi
c
h
is
c
om
pos
e
d
of
a
r
ti
c
le
s
f
r
om
th
e
bi
om
e
di
c
a
l
f
ie
ld
.
T
he
s
e
nt
e
nc
e
pa
ir
s
i
nc
lu
de
d i
n t
he
da
ta
s
e
t
w
e
r
e
c
hos
e
n f
r
om
s
e
nt
e
nc
e
s
t
ha
t
c
i
te
a
r
e
f
e
r
e
nc
e
a
r
ti
c
le
.
4.
M
E
T
H
O
D
4.1. P
r
e
t
r
ai
n
in
g t
h
e
B
E
R
T
m
od
e
l
T
he
or
ig
in
a
l
B
E
R
T
(
la
r
ge
)
w
e
ig
ht
s
tr
a
in
e
d
on
th
e
ge
ne
r
a
l
dom
a
in
,
na
m
e
ly
B
ooks
C
or
pus
a
nd
W
ik
ip
e
di
a
,
w
e
r
e
e
m
pl
oye
d
a
s
a
s
ta
r
ti
ng
poi
nt
[
1]
.
T
o
a
da
pt
B
E
R
T
f
or
bi
om
e
di
c
a
l
te
xt
,
pr
e
tr
a
in
in
g
w
a
s
pe
r
f
or
m
e
d
on
a
s
ubs
ta
nt
ia
l
vol
um
e
of
P
ubM
e
d
a
nd
M
I
M
I
C
-
I
I
I
a
nd
c
li
ni
c
a
l
not
e
s
.
T
he
pr
e
-
tr
a
in
in
g
a
nd
f
in
e
-
tu
ni
ng pr
oc
e
s
s
e
s
a
r
e
br
ie
f
ly
de
pi
c
te
d i
n F
ig
ur
e
1
.
F
ig
ur
e
1
. O
ve
r
vi
e
w
of
t
he
pr
e
tr
a
in
in
g a
nd f
in
e
-
tu
ni
ng s
ta
ge
s
of
M
e
di
c
a
lB
E
R
T
T
he
s
t
a
ti
s
ti
c
s
of
th
e
pr
e
tr
a
in
in
g
c
or
por
a
a
r
e
gi
ve
n
in
T
a
bl
e
2.
T
he
W
or
dP
ie
c
e
to
ke
ni
z
a
ti
on
te
c
hni
que
[
34]
w
hi
c
h
w
a
s
e
m
pl
oye
d
by
th
e
or
ig
in
a
l
B
E
R
T
,
is
a
s
ub
w
or
d
to
ke
ni
z
a
ti
on
m
e
th
od,
is
w
e
ll
s
ui
te
d
f
or
bi
om
e
di
c
a
l
c
or
por
a
be
c
a
us
e
it
c
a
n
e
f
f
e
c
ti
ve
ly
m
a
na
ge
va
r
io
u
s
ty
pe
s
of
voc
a
bul
a
r
y,
in
c
lu
di
ng
s
pe
c
ia
li
z
e
d
m
e
di
c
a
l
te
r
m
in
ol
ogy,
a
c
r
onyms
,
a
nd
r
a
r
e
or
out
-
of
-
voc
a
bul
a
r
y
w
or
ds
c
om
m
onl
y
f
ound
in
bi
om
e
di
c
a
l
te
xt
s
.
B
io
B
E
R
T
[
4]
f
or
to
ke
ni
z
a
ti
on.
B
ut
it
w
a
s
obs
e
r
ve
d
th
a
t
us
in
g
dom
a
in
-
s
pe
c
if
ic
voc
a
bul
a
r
y,
a
s
e
xpe
r
im
e
nt
e
d
in
th
e
w
or
k
by
[
35
]
,
ga
ve
be
tt
e
r
r
e
s
ul
ts
,
a
nd
in
th
e
w
or
k
c
ondu
c
te
d
by
F
a
c
e
book
r
e
s
e
a
r
c
h
on
R
oB
E
R
T
a
[
36]
,
in
c
r
e
a
s
e
d t
he
pe
r
f
or
m
a
nc
e
of
th
e
m
ode
l
f
o
r
dom
a
in
-
s
pe
c
if
ic
ta
s
ks
. H
e
nc
e
,
c
us
to
m
bi
om
e
di
c
a
l
voc
a
bul
a
r
y w
a
s
us
e
d
th
a
t
f
ol
lo
w
s
byt
e
-
pa
ir
e
nc
odi
ng
(
B
P
E
)
[
37]
,
[
38]
le
a
r
ne
d
f
r
om
P
ubM
e
d
pr
e
pr
oc
e
s
s
e
d
te
xt
[
39]
,
a
nd
a
li
gne
d
it
w
it
h
th
e
B
E
R
T
to
ke
ni
z
e
r
s
o
th
a
t
it
c
or
r
e
s
ponde
d
to
th
e
e
m
be
ddi
ngs
in
th
e
m
ode
l.
F
or
th
e
e
xpe
r
im
e
nt
a
l
s
e
tu
p,
th
e
m
a
xi
m
um
s
e
que
n
c
e
le
ngt
h
w
a
s
s
e
t
t
o
512,
a
nd
th
e
s
a
m
e
ba
tc
h
s
iz
e
w
a
s
u
s
e
d
f
or
pr
e
tr
a
in
in
g
B
E
R
T
on
g
e
ne
r
a
l
c
or
por
a
[
1]
.
T
he
l
e
a
r
ni
ng
r
a
te
w
a
s
s
e
t
to
3
e
-
4,
a
nd
th
e
m
ode
l
w
a
s
tr
a
in
e
d
f
or
450k
s
te
ps
on
P
ubM
e
d
a
nd
200k
s
te
ps
on
M
I
M
I
C
-
III
[
40]
on
N
V
I
D
I
A
A
100
G
P
U
s
w
it
h
to
ta
l
w
a
r
m
up
s
te
ps
of
2,000.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
e
di
c
al
B
E
R
T
:
e
nhanc
in
g bi
om
e
di
c
al
nat
ur
al
l
anguage
pr
oc
e
s
s
in
g us
in
g pr
e
tr
ai
ne
d
…
(
K
. Sahit
R
e
ddy
)
2373
T
a
bl
e
2. S
ta
ti
s
ti
c
s
of
pr
e
tr
a
in
in
g c
or
por
a
C
or
pus
S
t
a
t
i
s
t
i
c
s
D
om
a
i
n
E
ngl
i
s
h W
i
ki
pe
di
a
>5M
a
r
t
i
c
l
e
s
E
ngl
i
s
h (
G
e
ne
r
a
l
)
>3B
w
or
ds
B
ooks
C
or
pus
>11,038 books
G
e
ne
r
a
l
>985M
w
or
ds
P
ubM
e
d A
bs
t
r
a
c
t
s
>22M
a
bs
t
r
a
c
t
s
B
i
om
e
di
c
a
l
>5B
w
or
ds
P
ubM
e
d C
e
nt
r
a
l
>3M
a
r
t
i
c
l
e
s
B
i
om
e
di
c
a
l
>9.6B
w
or
ds
M
I
M
I
C
-
III
>2M
he
a
l
t
h
-
r
e
l
a
t
e
d not
e
s
M
e
di
c
a
l
>0.2B
w
or
ds
4.2. F
in
e
t
u
n
i
n
g t
h
e
B
E
R
T
m
od
e
l
F
in
e
-
tu
ni
ng
in
vol
ve
s
f
ur
th
e
r
tr
a
in
in
g
th
e
p
r
e
tr
a
in
e
d
m
ode
l
on
a
s
pe
c
if
ic
ta
s
k
or
dom
a
in
to
e
nha
nc
e
it
s
pe
r
f
or
m
a
nc
e
f
or
th
a
t
pa
r
ti
c
ul
a
r
a
ppl
ic
a
ti
on.
T
h
e
goa
l
w
a
s
to
f
in
e
-
tu
ne
th
e
B
E
R
T
m
ode
l
on
s
ix
be
nc
hm
a
r
k
ta
s
ks
a
v
a
il
a
bl
e
in
B
L
U
R
B
[
20]
a
nd
B
io
B
E
R
T
[
4]
r
e
pos
it
or
y.
W
e
us
e
d
f
our
pr
e
pr
oc
e
s
s
e
d
da
ta
s
e
ts
,
i.
e
.,
B
C
4C
H
E
M
D
,
L
I
N
N
A
E
U
S
,
a
nd
E
U
-
A
D
R
, a
va
il
a
bl
e
in
th
e
B
io
B
E
R
T
r
e
pos
it
or
y. T
he
bi
om
e
di
c
a
l
a
nd
c
li
ni
c
a
l
da
ta
s
e
ts
w
e
r
e
ut
il
iz
e
d
f
or
uns
upe
r
vi
s
e
d
f
in
e
-
tu
ni
ng,
a
da
pt
in
g
th
e
m
ode
l'
s
r
e
pr
e
s
e
nt
a
ti
ons
to
pe
r
f
or
m
e
f
f
e
c
ti
ve
ly
on va
r
io
us
bi
om
e
di
c
a
l
ta
s
k
s
.
W
e
us
e
d t
w
o N
V
I
D
I
A
A
100 G
P
U
s
w
it
h pa
r
a
ll
e
l
t
r
a
in
in
g t
o
f
in
e
-
tu
ne
our
B
E
R
T
m
ode
l
on e
a
c
h t
a
s
k.
F
P
16
m
ix
e
d
-
p
r
e
c
is
io
n
w
a
s
a
ls
o
us
e
d
to
s
pe
e
d
up
tr
a
in
in
g
a
nd
r
e
duc
e
m
e
m
or
y
us
a
ge
.
I
t
to
ok
a
ppr
oxi
m
a
te
ly
15
-
20
m
in
ut
e
s
to
f
in
e
-
tu
ne
th
e
m
ode
l
f
o
r
s
e
que
nc
e
la
be
li
ng
a
n
d
s
e
que
nc
e
c
l
a
s
s
if
ic
a
ti
on
ta
s
ks
.
F
or
s
e
que
nc
e
la
be
li
ng, ba
tc
h s
iz
e
s
of
16 a
nd 32 we
r
e
us
e
d, a
nd t
he
l
e
a
r
ni
ng r
a
te
w
a
s
5e
-
5. F
or
t
he
J
N
L
P
B
A
[
28]
c
or
pus
, w
e
s
e
t
a
le
a
r
ni
ng
r
a
te
of
1e
-
5.
F
or
da
ta
s
e
t
s
w
it
h
hi
ghe
r
num
b
e
r
of
e
xa
m
pl
e
s
,
a
tr
a
in
/v
a
li
da
ti
on/
te
s
t
s
pl
it
of
a
bout
80/
10/
10
or
75/
15/
10
is
m
a
de
,
w
he
r
e
a
s
f
or
s
m
a
ll
e
r
da
ta
s
e
ts
,
t
he
s
pl
it
is
a
bout
60/
20/
20.
T
h
e
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
da
ta
s
e
ts
w
e
r
e
tr
a
in
e
d
b
e
twe
e
n
5
a
nd
25
e
poc
hs
.
O
n
th
e
ot
he
r
ha
nd,
f
or
s
e
qu
e
nc
e
c
la
s
s
if
ic
a
ti
on,
a
ba
tc
h
s
iz
e
of
32
a
nd
le
a
r
ni
ng
r
a
te
s
of
2e
-
5
a
nd
3e
-
5
w
e
r
e
us
e
d
.
T
he
m
ode
l
w
a
s
tr
a
in
e
d
f
or
10
e
poc
hs
on
th
e
C
he
m
P
r
ot
, D
D
I
-
2013 a
nd G
A
D
da
ta
s
e
ts
.
5.
R
E
S
U
L
T
S
A
N
D
A
N
A
L
Y
S
I
S
T
he
r
e
s
ul
t
s
of
M
e
di
c
a
lB
E
R
T
w
e
r
e
c
om
pa
r
e
d
w
it
h
th
r
e
e
s
pe
c
ia
li
z
e
d
B
E
R
T
m
ode
ls
pr
e
s
e
nt
e
d
in
T
a
bl
e
3.
T
h
e
s
e
m
ode
l
s
in
c
lu
de
S
c
iB
E
R
T
[
8]
,
C
li
ni
c
a
lB
E
R
T
[
7]
,
a
nd
B
io
B
E
R
T
[
4]
.
W
e
a
ls
o
c
om
pa
r
e
d
th
e
r
e
s
ul
ts
w
it
h
th
e
or
ig
in
a
l
B
E
R
T
m
ode
l
(
c
ol
um
n
3)
tr
a
in
e
d
on
ge
ne
r
a
l
E
ngl
is
h
c
or
por
a
a
nd
R
oB
E
R
T
a
[
9]
(
c
ol
um
n
4)
to
e
va
lu
a
te
th
e
pe
r
f
or
m
a
nc
e
of
th
e
m
ode
l
on
N
L
P
ta
s
ks
a
nd
it
s
unde
r
s
ta
ndi
ng
in
bi
om
e
di
c
a
l,
c
li
ni
c
a
l
a
nd
s
c
ie
nt
if
ic
te
r
m
s
.
T
he
r
e
s
ul
ts
of
M
e
di
c
a
lB
E
R
T
unde
r
th
e
va
r
io
us
e
va
lu
a
ti
on
m
e
t
r
ic
s
on
s
e
que
nc
e
la
be
li
ng,
r
e
la
ti
on
e
xt
r
a
c
ti
on
a
nd
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on
a
r
e
s
how
n
in
T
a
bl
e
4.
O
n
s
e
que
nc
e
la
be
li
ng
t
a
s
ks
,
B
io
B
E
R
T
out
pe
r
f
or
m
e
d t
he
ot
he
r
m
ode
ls
on t
hr
e
e
da
ta
s
e
ts
i
n t
he
bi
om
e
di
c
a
l
dom
a
in
. T
he
pe
r
f
or
m
a
nc
e
of
t
he
R
oB
E
R
T
a
m
ode
l
is
c
om
pa
r
a
bl
e
to
th
a
t
of
th
e
pr
opos
e
d
m
ode
l
on
L
I
N
N
A
E
U
S
[
30]
,
w
it
h
a
n
F
1
s
c
or
e
of
87.8.
M
e
di
c
a
lB
E
R
T
a
c
hi
e
ve
d
th
e
b
e
s
t
r
e
s
ul
t
s
on
f
iv
e
out
of
e
ig
ht
da
ta
s
e
ts
in
th
e
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
ta
s
k.
O
ve
r
a
ll
,
th
e
m
ode
l
out
p
e
r
f
or
m
s
R
oB
E
R
T
a
by
2.04
a
nd
c
om
pe
ti
ti
ve
ly
out
pe
r
f
or
m
s
B
io
B
E
R
T
by
0.75
(
m
e
a
n
F
1 i
nc
r
e
a
s
e
)
a
nd S
c
iB
E
R
T
by 1.4.
T
he
va
r
ia
ti
on
in
th
e
F
1
e
va
lu
a
ti
on
m
e
tr
ic
of
th
e
th
r
e
e
m
od
e
ls
a
nd
M
e
di
c
a
lB
E
R
T
a
c
r
os
s
th
e
f
our
s
e
que
nc
e
l
a
be
li
ng
ta
s
k
s
is
de
pi
c
t
e
d
in
F
ig
ur
e
2(
a
)
.
C
om
pa
r
in
g
w
it
h
B
io
B
E
R
T
,
M
e
di
c
a
lB
E
R
T
a
c
hi
e
v
e
d
be
tt
e
r
r
e
s
ul
ts
in
r
e
la
ti
on
e
xt
r
a
c
ti
on
a
nd
s
e
que
n
c
e
c
la
s
s
if
ic
a
ti
on
ta
s
ks
w
it
h
s
im
il
a
r
s
c
or
e
s
on
two
da
ta
s
e
ts
(
85.6 a
nd 85.5 f
or
G
A
D
a
nd 76.8 a
nd 76.1
f
or
D
D
I
-
2013
)
.
T
he
m
ode
l’
s
pe
r
f
or
m
a
nc
e
i
s
s
im
il
a
r
ly
c
om
pa
r
e
d t
o
ot
he
r
s
a
nd
de
pi
c
te
d
in
F
ig
ur
e
2(
b)
.
M
e
di
c
a
l
B
E
R
T
f
a
c
e
d
c
om
pe
ti
ti
on
w
it
h
S
c
iB
E
R
T
on
E
U
-
A
D
R
da
ta
s
e
t,
w
it
h
m
ic
r
o
a
ve
r
a
ge
F
1
s
c
or
e
s
of
85.6
a
nd
81.2
r
e
s
pe
c
ti
v
e
ly
.
T
he
pe
r
f
or
m
a
nc
e
of
our
pr
opos
e
d
m
ode
l
on
s
e
nt
e
nc
e
s
im
il
a
r
it
y,
que
s
ti
on
a
n
s
w
e
r
in
g
a
nd
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on
is
de
pi
c
te
d
in
F
ig
ur
e
2(
c
)
.
T
he
be
tt
e
r
r
e
s
ul
ts
of
M
e
di
c
a
lB
E
R
T
c
om
pa
r
e
d
to
th
e
ot
he
r
m
ode
ls
f
or
m
a
ny
be
nc
hm
a
r
k
da
ta
s
e
ts
r
e
por
te
d
c
a
n
be
a
tt
r
ib
ut
e
d
to
la
r
ge
r
a
nd
r
e
le
va
nt
pr
e
-
tr
a
in
in
g
da
ta
,
a
s
c
om
pa
r
e
d
to
C
li
ni
c
a
lB
E
R
T
(
w
hi
c
h
is
tr
a
in
e
d
on
m
a
in
ly
c
li
ni
c
a
l
doc
um
e
nt
s
)
,
S
c
i
B
E
R
T
(
w
hi
c
h
is
tr
a
in
e
d
on
s
c
ie
nt
if
i
c
doc
um
e
nt
s
a
nd
l
e
s
s
bi
om
e
di
c
a
l
te
xt
)
a
nd
R
oB
E
R
T
a
(
w
hi
c
h
is
tr
a
in
e
d
on
ge
ne
r
a
l
pur
pos
e
E
ngl
is
h
te
xt
)
,
a
nd
dom
a
in
-
s
pe
c
if
ic
bi
om
e
di
c
a
l
voc
a
bul
a
r
y
us
e
d
c
om
pa
r
e
d
to
th
e
ot
he
r
m
ode
ls
.
T
he
tr
a
de
of
f
c
a
us
e
d
due
t
o
c
om
put
a
ti
ona
l
c
os
t
to
tr
a
in
th
e
m
ode
l
on
a
la
r
ge
r
bi
om
e
di
c
a
l
a
nd
c
li
ni
c
a
l
c
or
pus
f
ound
th
e
ne
e
d
to
r
e
du
c
e
th
e
num
be
r
of
tr
a
in
in
g
s
te
ps
c
om
pa
r
e
d
to
ot
he
r
m
ode
ls
in
th
e
pr
e
-
tr
a
in
in
g
s
ta
ge
,
a
nd
di
f
f
e
r
e
nc
e
s
in
hyp
e
r
pa
r
a
m
e
te
r
s
s
e
t
dur
in
g
th
e
f
in
e
-
tu
ni
ng
s
ta
ge
w
it
h
th
e
ot
he
r
s
m
a
ke
th
e
m
pos
s
ib
le
r
e
a
s
ons
f
or
poor
e
r
r
e
s
ul
ts
in
s
om
e
ta
s
ks
.
A
ls
o,
th
e
una
va
il
a
bi
li
ty
of
pr
ope
r
a
nd
hi
gh
s
ta
nda
r
d
d
a
ta
s
e
t
s
f
or
f
in
e
-
tu
ni
ng
f
or
s
e
que
nc
e
la
be
li
ng
ta
s
k,
r
e
la
ti
on
e
xt
r
a
c
ti
on
a
nd
s
e
nt
e
nc
e
s
im
il
a
r
it
y t
a
s
ks
a
ls
o r
e
s
ul
te
d i
n a
de
c
r
e
a
s
e
i
n pe
r
f
or
m
a
nc
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2367
-
2378
2374
T
a
bl
e
3. P
e
r
f
or
m
a
nc
e
of
B
E
R
T
a
nd othe
r
m
ode
l
s
on 17 be
nc
hm
a
r
k da
ta
s
e
ts
D
a
t
a
s
e
t
M
e
t
r
i
c
B
E
R
T
R
oB
E
R
T
a
S
c
i
B
E
R
T
C
l
i
ni
c
a
l
B
E
R
T
B
i
oB
E
R
T
M
e
di
c
a
l
B
E
R
T
(
o
ur
s
)
N
C
B
I
-
D
i
s
e
a
s
e
F1
85.6
86.6
88.2
86.3
89.1
88.3*
B
C
2G
M
F1
81.8
80.9
83.4*
81.7
83.8
82.8
B
C
5C
D
R
-
C
h
e
m
F1
91.2
90.8
92.5
90.8
92.8*
93.2
B
C
5C
D
R
-
D
i
s
e
a
s
e
F1
82.4
82.3
84.5
83.0
84.7*
86.8
J
N
L
P
B
A
F1
90.0
90.6
91.8*
90.3
92.2
91.6
L
I
N
N
A
E
U
S
F1
84.3
87.8
84.1
84.8
86.2*
87.8
B
C
4C
H
E
M
D
F1
74.9
80.1
79.6
78.6
80.4*
83.6
S
pe
c
i
e
s
-
800
M
a
c
r
o
-
F1
72.3
73.0
73.1
72.1
73.2*
74.3
E
B
M
-
P
I
C
O
M
i
c
r
o
-
F1
77.7
77.7
80.9*
78.4
80.9*
85.8
GAD
F1
84.6
85.0
85.5*
85.1
85.0
85.6
EU
-
ADR
M
i
c
r
o
-
F1
80.0
79.5
81.2*
78.2
80.9
85.4
DDI
-
2013
M
i
c
r
o
-
F1
71.9
72.9
75.2
72.0
76.1*
76.8
C
he
m
P
r
ot
A
c
c
ur
a
c
y
82.7
81.2
87.1
91.2*
89.5
92.0
P
ubM
e
dQ
A
A
c
c
ur
a
c
y
51.6
52.8
57.4
49.1
60.2*
61.3
B
i
oA
S
Q
P
e
a
r
s
on
74.4
75.2
78.9
68.5
84.1*
87.9
B
I
O
S
S
E
S
M
i
c
r
o
-
F1
80.2
79.6
81.1
80.7
81.5*
81.6
H
oC
F1
85.6
86.6
88.2
86.3
89.1
88.3*
T
ot
a
l
(
M
e
a
n)
79.5
80.2
81.9
79.8
82.6*
84.3
B
ol
d
s
c
or
e
s
de
not
e
t
he
be
s
t
, a
s
t
e
r
i
s
k (
*)
m
a
r
ks
t
he
s
e
c
ond
-
be
s
t
, a
nd unde
r
l
i
ne
d
s
c
or
e
s
i
ndi
c
a
t
e
t
i
e
s
T
a
bl
e
4. M
e
di
c
a
lB
E
R
T
r
e
s
ul
ts
on na
m
e
d e
nt
it
y r
e
c
ogni
ti
on, r
e
la
ti
on e
xt
r
a
c
ti
on, a
nd doc
um
e
nt
c
la
s
s
if
ic
a
ti
on
e
va
lu
a
te
d by F
1, r
e
c
a
ll
, a
nd pr
e
c
is
io
n
D
a
t
a
s
e
t
F1
R
(
R
e
c
a
l
l
)
P
(
P
r
e
c
i
s
i
on)
N
C
B
I
–
D
i
s
e
a
s
e
88.301
90.417
86.282
B
C
2G
M
82.754
83.304
82.212
B
C
5C
D
R
–
C
he
m
93.162
93.110
92.214
B
C
5C
D
R
-
D
i
s
e
a
s
e
86.784
86.618
84.966
L
I
N
N
A
E
U
S
87.822
84.042
85.365
B
C
4C
H
E
M
D
83.566
84.355
82.791
J
N
L
P
B
A
91.651
89.983
92.328
S
pe
c
i
e
s
–
800
74.315
72.565
70.895
DDI
–
2013
76.889
75.485
78.367
EU
–
ADR
85.433
86.150
74.206
C
he
m
P
r
ot
92.053
90.053
88.622
GAD
86.604
92.833
79.765
H
oC
88.378
81.311
84.678
B
ol
d i
ndi
c
a
t
e
s
be
s
t
s
c
or
e
s
M
e
di
c
a
lB
E
R
T
’
s
pr
a
c
t
ic
a
l
a
p
pl
i
c
a
ti
on
s
i
n
h
e
a
lt
h
c
a
r
e
a
nd
b
i
om
e
di
c
a
l
f
i
e
ld
s
a
r
e
tr
a
n
s
f
or
m
a
ti
ve
,
e
s
p
e
c
i
a
l
ly
a
s
th
e
m
od
e
l
a
d
a
pt
s
t
o
t
he
c
o
m
pl
e
x nu
a
n
c
e
s
of
m
e
di
c
a
l
l
a
n
gu
a
g
e
. B
y
f
in
e
-
t
uni
ng B
E
R
T
t
o
r
e
c
og
ni
z
e
a
nd
in
te
r
pr
e
t
b
io
m
e
d
ic
a
l
te
xt
,
M
e
di
c
a
l
B
E
R
T
e
f
f
e
c
t
iv
e
ly
a
ddr
e
s
s
e
s
s
e
ve
r
a
l
c
r
i
ti
c
a
l
n
e
e
d
s
a
c
r
os
s
v
a
r
io
u
s
d
om
a
in
s
in
h
e
a
lt
h
c
a
r
e
,
e
nh
a
n
c
in
g
d
e
c
i
s
io
n
s
up
por
t
,
di
a
g
no
s
t
ic
a
c
c
u
r
a
c
y
,
a
nd
r
e
s
e
a
r
c
h
a
dv
a
n
c
e
m
e
nt
s
. W
it
h
th
e
c
a
pa
bi
l
it
y
to
c
a
pt
ur
e
c
o
nt
e
xt
f
r
om
bi
om
e
di
c
a
l
t
e
x
t,
M
e
d
ic
a
l
B
E
R
T
f
a
c
i
li
t
a
t
e
s
e
f
f
ic
ie
nt
e
x
tr
a
c
t
io
n
of
c
r
i
ti
c
a
l
in
f
or
m
a
ti
on
f
r
om
l
a
r
g
e
vo
lu
m
e
s
of
m
e
d
ic
a
l
do
c
um
e
n
ts
.
T
hi
s
i
nc
lu
d
e
s
i
de
nt
if
yi
ng
a
n
d
c
a
t
e
g
or
i
z
in
g
s
ym
pt
o
m
s
,
m
e
di
c
a
ti
on
s
,
tr
e
a
tm
e
n
t
pl
a
ns
,
a
nd
l
a
b
r
e
s
ul
t
s
f
r
o
m
f
r
e
e
-
te
xt
E
H
R
s
.
M
e
di
c
a
l
in
f
or
m
a
ti
o
n
e
xt
r
a
c
ti
on
not
on
ly
a
s
s
i
s
t
s
i
n
p
a
t
ie
nt
c
a
r
e
bu
t
a
l
s
o
a
c
c
e
le
r
a
t
e
s
r
e
s
e
a
r
c
h
a
c
t
iv
it
ie
s
b
y
s
tr
uc
tu
r
i
ng
d
a
t
a
f
or
u
s
e
in
la
r
g
e
-
s
c
a
l
e
s
t
udi
e
s
a
n
d
m
e
t
a
-
a
n
a
l
ys
e
s
.
M
e
di
c
a
lB
E
R
T
c
a
n
c
o
nt
r
i
but
e
to
p
a
t
ie
nt
s
a
f
e
t
y
by
id
e
nt
if
yi
n
g
p
ot
e
nt
i
a
l
dr
ug
-
dr
u
g
in
t
e
r
a
c
ti
o
ns
.
B
y
tr
a
in
i
ng
o
n
la
r
g
e
d
a
t
a
s
e
t
s
of
d
r
ug
i
nt
e
r
a
c
t
io
n
s
a
nd
E
H
R
s
,
M
e
di
c
a
lB
E
R
T
c
a
n
a
na
ly
z
e
pr
e
s
c
r
ip
ti
o
n
s
a
n
d
a
l
e
r
t
c
li
ni
c
ia
n
s
to
da
n
ge
r
ou
s
c
om
bi
n
a
t
io
n
s
,
th
u
s
r
e
d
uc
in
g
t
he
r
is
k
of
a
dv
e
r
s
e
e
v
e
n
t
s
in
p
ol
yph
a
r
m
a
c
y
s
c
e
n
a
r
i
o
s
.
F
u
tu
r
e
w
or
k
m
a
y
e
xp
lo
r
e
f
ur
th
e
r
c
u
s
to
m
iz
a
t
io
n
a
n
d
e
v
a
lu
a
ti
on
of
M
e
d
ic
a
l
B
E
R
T
a
c
r
o
s
s
t
he
s
e
a
ppl
ic
a
t
io
n
s
t
o
o
pt
im
iz
e
i
ts
in
t
e
gr
a
ti
on
i
nt
o r
e
a
l
-
w
or
l
d
he
a
lt
hc
a
r
e
s
y
s
te
m
s
a
nd
c
li
ni
c
a
l
w
or
kf
l
ow
s
.
T
h
e
u
s
e
of
M
e
di
c
a
lB
E
R
T
a
nd
ot
he
r
tr
a
n
s
f
or
m
e
r
-
b
a
s
e
d
m
od
e
l
s
in
t
h
e
b
io
m
e
d
ic
a
l
f
i
e
l
d
o
p
e
n
s
up
v
a
s
t
po
s
s
ib
i
li
t
ie
s
f
or
im
pr
ovi
ng
he
a
lt
hc
a
r
e
a
nd
c
li
ni
c
a
l
r
e
s
e
a
r
c
h
.
H
ow
e
v
e
r
,
t
he
s
e
n
s
it
iv
e
n
a
tu
r
e
of
m
e
di
c
a
l
da
ta
in
tr
o
du
c
e
s
c
om
pl
e
x
e
th
i
c
a
l
a
nd pr
i
va
c
y c
o
nc
e
r
ns
. B
io
m
e
di
c
a
l
d
a
ta
i
s
i
nh
e
r
e
nt
ly
s
e
n
s
it
iv
e
, c
o
nt
a
in
in
g pe
r
s
on
a
ll
y
id
e
nt
if
ia
bl
e
in
f
or
m
a
ti
on
(
P
I
I
)
a
n
d
pr
ot
e
c
t
e
d
he
a
l
th
i
nf
or
m
a
t
io
n
(
P
H
I
)
t
ha
t,
if
e
x
po
s
e
d,
c
o
ul
d
ha
ve
s
e
v
e
r
e
c
on
s
e
q
ue
nc
e
s
f
or
in
d
iv
i
du
a
l
s
.
G
iv
e
n
M
e
d
ic
a
l
B
E
R
T
’
s
r
e
li
a
n
c
e
on
l
a
r
ge
d
a
t
a
s
e
t
s
f
or
e
f
f
e
c
ti
ve
pr
e
-
tr
a
i
ni
n
g
a
n
d
f
in
e
-
tu
ni
n
g,
e
n
s
u
r
in
g d
a
t
a
pr
iv
a
c
y d
ur
in
g m
od
e
l
de
ve
lo
p
m
e
nt
i
s
e
s
s
e
n
ti
a
l.
T
hi
s
c
a
n b
e
d
on
e
b
y
e
s
ta
bl
i
s
hi
n
g c
le
a
r
gui
d
e
l
in
e
s
f
or
da
ta
a
no
nym
iz
a
t
io
n
,
e
n
c
r
ypt
io
n
,
a
n
d
di
f
f
e
r
e
nt
ia
l
pr
iv
a
c
y
te
c
h
ni
qu
e
s
i
n
m
e
di
c
a
l
A
I
a
p
pl
i
c
a
ti
on
s
.
T
h
e
u
s
e
of
m
e
di
c
a
l
r
e
c
or
ds
f
or
tr
a
i
ni
n
g
m
od
e
l
s
li
k
e
M
e
d
ic
a
l
B
E
R
T
s
ho
ul
d
b
e
a
c
c
om
p
a
ni
e
d
by
tr
a
n
s
p
a
r
e
nt
c
on
s
e
n
t
pr
a
c
ti
c
e
s
.
P
a
ti
e
nt
s
s
h
oul
d
b
e
in
f
or
m
e
d
a
b
out
ho
w
th
e
ir
da
ta
m
a
y
c
ont
r
ib
ut
e
to
t
r
a
i
ni
ng
m
od
e
l
s
f
or
va
r
i
ou
s
a
pp
li
c
a
ti
o
n
s
,
a
lo
ng
w
it
h
th
e
p
ot
e
nt
i
a
l
be
ne
f
it
s
a
nd
r
is
k
s
.
P
ol
i
c
i
e
s
r
e
qui
r
in
g
r
e
gul
a
r
e
th
ic
a
l
a
n
d
f
a
i
r
n
e
s
s
a
ud
it
s
c
a
n
e
n
s
ur
e
t
ha
t
m
od
e
l
s
l
ik
e
M
e
di
c
a
lB
E
R
T
m
e
e
t
hi
g
h
s
t
a
nda
r
d
s
of
tr
a
n
s
p
a
r
e
n
c
y
a
n
d
a
c
c
o
unt
a
b
il
i
ty
.
A
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
e
di
c
al
B
E
R
T
:
e
nhanc
in
g bi
om
e
di
c
al
nat
ur
al
l
anguage
pr
oc
e
s
s
in
g us
in
g pr
e
tr
ai
ne
d
…
(
K
. Sahit
R
e
ddy
)
2375
M
e
di
c
a
lB
E
R
T
be
c
o
m
e
s
i
nt
e
gr
a
t
e
d
i
nt
o
c
l
in
i
c
a
l
w
or
kf
lo
w
s
,
th
e
ne
e
d
f
or
e
xp
la
in
a
bi
li
t
y
gr
ow
s
.
P
ol
i
c
y
f
r
a
m
e
w
or
k
s
t
ha
t
m
a
n
da
te
e
x
pl
a
in
a
bi
li
ty
in
A
I
-
dr
iv
e
n
de
c
i
s
i
on
-
m
a
ki
ng
pr
o
c
e
s
s
e
s
c
a
n
pr
om
o
te
a
c
c
oun
ta
bi
l
it
y
a
nd
in
c
r
e
a
s
e
m
o
de
l
a
c
c
e
pt
a
n
c
e
a
m
o
ng
c
l
in
i
c
i
a
n
s
, i
m
pr
ovi
ng
ov
e
r
a
ll
t
r
us
t
i
n t
he
t
e
c
hn
ol
o
gy.
(
a
)
(
b)
(
c
)
F
ig
ur
e
2.
P
e
r
f
or
m
a
nc
e
va
r
ia
ti
on of
bi
om
e
di
c
a
l
m
ode
ls
a
nd pr
opos
e
d m
ode
l
on (
a
)
s
e
que
nc
e
l
a
be
li
ng
;
(
b)
r
e
la
ti
on e
xt
r
a
c
ti
on
;
a
nd (
c
)
s
e
nt
e
nc
e
s
im
il
a
r
it
y, que
s
ti
on a
ns
w
e
r
in
g a
nd doc
um
e
nt
c
la
s
s
if
ic
a
ti
on t
a
s
ks
6.
C
O
N
C
L
U
S
I
O
N
T
he
B
E
R
T
m
ode
l
ha
s
be
e
n
s
ubj
e
c
te
d
to
s
ix
va
r
io
us
be
nc
hm
a
r
k
ta
s
ks
on
17
da
ta
s
e
ts
.
T
he
s
e
ta
s
ks
w
e
r
e
a
ddr
e
s
s
e
d
by
u
s
in
g
tr
a
ns
f
e
r
le
a
r
ni
ng
te
c
hni
que
s
.
T
hi
s
m
ode
l
out
pe
r
f
or
m
s
th
e
ot
he
r
s
pe
c
ia
li
z
e
d
B
E
R
T
m
ode
ls
on
s
e
que
n
c
e
c
l
a
s
s
if
ic
a
ti
on,
s
e
nt
e
n
c
e
s
im
il
a
r
it
y
ta
s
k
s
,
q
ue
s
ti
on
a
n
s
w
e
r
in
g
a
nd
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on
ta
s
ks
.
W
e
pl
a
n
to
tr
a
in
th
e
B
E
R
T
-
ba
s
e
d
va
r
ia
nt
on
la
r
ge
r
bi
om
e
di
c
a
l
c
or
por
a
a
nd
c
li
ni
c
a
l
r
e
por
ts
u
s
in
g
be
tt
e
r
c
om
put
a
ti
ona
l
r
e
s
our
c
e
s
,
to
d
e
te
r
m
in
e
th
e
e
f
f
e
c
t
of
tr
a
in
in
g
c
or
por
a
a
nd
th
e
ir
le
ngt
h
on
pe
r
f
or
m
a
nc
e
in
th
e
f
ut
ur
e
.
A
ddi
ti
ona
ll
y,
w
e
a
ls
o
pl
a
n
to
e
va
lu
a
te
M
e
di
c
a
lB
E
R
T
on
th
e
f
iv
e
bi
om
e
di
c
a
l
te
xt
-
m
in
in
g
ta
s
ks
w
it
h
te
n
c
or
por
a
f
r
om
th
e
b
io
m
e
di
c
a
l
la
ngua
ge
unde
r
s
ta
ndi
ng
e
va
l
ua
ti
on
(
B
L
U
E
)
be
nc
hm
a
r
k.
F
ur
th
e
r
m
o
r
e
,
th
e
f
ut
ur
e
w
or
k
a
ls
o
in
c
lu
de
s
li
ngui
s
ti
c
s
uppor
t
by
in
c
or
por
a
ti
ng
tr
a
in
in
g
a
nd
e
va
lu
a
ti
on
on
va
r
io
us
d
a
ta
s
e
t
s
a
nd
te
xt
s
pr
e
s
e
nt
e
d i
n l
a
ngua
g
e
s
ot
he
r
t
ha
n E
ngl
is
h t
o s
tu
dy t
he
i
m
p
a
c
t
on t
he
pe
r
f
or
m
a
nc
e
of
t
he
m
ode
l.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
W
e
a
c
knowle
dge
th
e
s
uppor
t
of
R
.
V
.
C
ol
le
ge
of
E
ngi
ne
e
r
i
ng
f
or
pr
ovi
d
in
g
th
e
c
om
put
a
ti
ona
l
r
e
s
our
c
e
s
ne
c
e
s
s
a
r
y t
o c
om
pl
e
te
t
hi
s
r
e
s
e
a
r
c
h.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2367
-
2378
2376
F
U
N
D
I
N
G
I
N
F
O
R
M
A
T
I
O
N
A
ut
hor
s
s
ta
te
no f
undi
ng i
nvol
ve
d.
A
U
T
H
O
R
C
O
N
T
R
I
B
U
T
I
O
N
S
S
T
A
T
E
M
E
N
T
T
hi
s
jo
ur
na
l
us
e
s
th
e
C
ont
r
ib
ut
or
R
ol
e
s
T
a
xonomy
(
C
R
e
di
T
)
to
r
e
c
ogni
z
e
in
di
vi
dua
l
a
ut
hor
c
ont
r
ib
ut
io
ns
, r
e
duc
e
a
ut
hor
s
hi
p di
s
put
e
s
,
a
nd f
a
c
il
it
a
te
c
ol
la
bo
r
a
ti
on.
N
am
e
o
f
A
u
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
K
. S
a
hi
t
R
e
ddy
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
N
. R
a
ga
ve
nde
r
a
n
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
V
a
s
a
nt
h K
.
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
G
a
ne
s
h N
. N
a
ik
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
V
is
ha
la
ks
hi
P
r
a
bhu
✓
✓
✓
✓
✓
✓
✓
N
a
ga
r
a
ja
G
. S
.
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
l
i
z
a
t
i
on
M
:
M
e
t
hodol
ogy
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
da
t
i
on
Fo
:
Fo
r
m
a
l
a
na
l
ys
i
s
I
:
I
nve
s
t
i
ga
t
i
on
R
:
R
e
s
our
c
e
s
D
:
D
a
t
a
C
ur
a
t
i
on
O
:
W
r
i
t
i
ng
-
O
r
i
gi
na
l
D
r
a
f
t
E
:
W
r
i
t
i
ng
-
R
e
vi
e
w
&
E
di
t
i
ng
Vi
:
Vi
s
ua
l
i
z
a
t
i
on
Su
:
Su
pe
r
vi
s
i
on
P
:
P
r
oj
e
c
t
a
dm
i
ni
s
t
r
a
t
i
on
Fu
:
Fu
ndi
ng a
c
qui
s
i
t
i
on
C
O
N
F
L
I
C
T
O
F
I
N
T
E
R
E
S
T
S
T
A
T
E
M
E
N
T
A
ut
hor
s
s
ta
te
no c
onf
li
c
t
of
i
nt
e
r
e
s
t.
D
A
T
A
A
V
A
I
L
A
B
I
L
I
T
Y
T
he
da
ta
th
a
t
s
uppor
t
th
e
f
in
di
ngs
of
th
is
s
tu
dy
w
il
l
be
a
va
il
a
bl
e
in
ht
tp
s
:/
/g
it
hub.c
om
/k
s
a
hi
tr
e
ddy/
M
e
di
c
a
lB
E
R
T
f
ol
lo
w
in
g
a
2
-
m
ont
h
e
m
ba
r
go
f
r
om
th
e
da
te
of
publ
ic
a
ti
on
to
a
ll
ow
f
or
t
he
c
om
m
e
r
c
ia
li
z
a
ti
on of
r
e
s
e
a
r
c
h f
in
di
ngs
.
R
E
F
E
R
E
N
C
E
S
[
1]
J
.
D
e
vl
i
n,
M
.
-
W
.
C
ha
ng,
K
.
L
e
e
,
K
.
T
.
G
oogl
e
,
a
nd
A
.
I
.
L
a
ngua
g
e
,
“
B
E
R
T
:
p
re
-
t
r
a
i
ni
ng
of
de
e
p
bi
di
r
e
c
t
i
ona
l
t
r
a
ns
f
or
m
e
r
s
f
or
l
a
ngua
ge
unde
r
s
t
a
ndi
ng,”
i
n
P
r
oc
e
e
di
ngs
of
N
A
A
C
L
-
H
L
T
2019
, pp. 4171
–
4186
, 2019
.
[
2]
Y
.
S
ha
o,
S
.
T
a
yl
o
r
,
N
.
M
a
r
s
ha
l
l
,
C
.
M
or
i
oka
,
a
nd
Q
.
Z
e
ng
-
T
r
e
i
t
l
e
r
,
“
C
l
i
ni
c
a
l
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on
w
i
t
h
w
or
d
e
m
be
ddi
ng
f
e
a
t
ur
e
s
vs
.
ba
g
-
of
-
w
or
ds
f
e
a
t
ur
e
s
,”
i
n
2018
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
B
i
g
D
at
a
(
B
i
g
D
at
a)
,
pp.
2874
–
2878
,
D
e
c
.
2018,
doi
:
10.1109/
B
i
gD
a
t
a
.2018.8622345.
[
3]
P
.
F
.
B
r
ow
n,
P
.
V
.
D
e
S
ouz
a
,
R
.
L
.
M
e
r
c
e
r
,
V
.
J
.
D
.
P
i
e
t
r
a
,
a
nd
J
.
C
.
L
a
i
,
“
C
l
a
s
s
-
ba
s
e
d
n
-
gr
a
m
m
ode
l
s
of
na
t
ur
a
l
l
a
ngua
ge
,
”
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
, vol
. 18, no. 4, pp. 467
–
479, 1992.
[
4]
J
.
L
e
e
e
t
al
.
,
“
B
i
oB
E
R
T
:
a
pr
e
-
t
r
a
i
ne
d
b
i
om
e
di
c
a
l
l
a
ngua
ge
r
e
pr
e
s
e
nt
a
t
i
on
m
ode
l
f
or
bi
om
e
di
c
a
l
t
e
xt
m
i
n
i
ng,”
B
i
oi
nf
or
m
at
i
c
s
,
vol
. 36, no. 4, pp. 1234
–
1240, F
e
b. 2020, doi
:
10.1093/
bi
oi
nf
or
m
a
t
i
c
s
/
bt
z
682.
[
5]
Y
.
P
e
ng,
S
.
Y
a
n,
a
nd
Z
.
L
u,
“
T
r
a
ns
f
e
r
l
e
a
r
ni
ng
i
n
bi
o
m
e
di
c
a
l
na
t
ur
a
l
l
a
ngua
ge
pr
oc
e
s
s
i
ng:
A
n
e
va
l
ua
t
i
on
of
B
E
R
T
a
nd
E
L
M
o
on
t
e
n
be
nc
hm
a
r
ki
ng
da
t
a
s
e
t
s
,”
B
i
oN
L
P
2019
-
SI
G
B
i
oM
e
d
W
or
k
s
hop
on
B
i
om
e
di
c
al
N
at
ur
al
L
anguage
P
r
oc
e
s
s
i
ng,
P
r
oc
e
e
di
ngs
o
f
t
he
18t
h B
i
oN
L
P
W
or
k
s
hop and Shar
e
d T
as
k
, pp. 58
–
65, 2019, doi
:
10.18653/
v
1/
w
19
-
5006.
[
6]
M
.
E
.
P
e
t
e
r
s
e
t
al
.
,
“
D
e
e
p
c
ont
e
xt
u
a
l
i
z
e
d
w
or
d
r
e
pr
e
s
e
nt
a
t
i
ons
,
”
N
A
A
C
L
H
L
T
2018
-
2018
C
onf
e
r
e
nc
e
of
t
he
N
o
r
t
h
A
m
e
r
i
c
a
n
C
hapt
e
r
of
t
he
A
s
s
oc
i
at
i
on f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
:
H
um
an L
anguage
T
e
c
hnol
ogi
e
s
-
P
r
oc
e
e
di
ngs
of
t
he
C
onf
e
r
e
nc
e
,
vol
. 1
,
pp. 2227
–
2237, 2018, doi
:
10.18653/
v1/
n18
-
1202.
[
7]
E
.
A
l
s
e
nt
z
e
r
e
t
al
.
,
“
P
ubl
i
c
l
y
a
va
i
l
a
bl
e
c
l
i
ni
c
a
l
B
E
R
T
e
m
be
ddi
ngs
,”
I
n
P
r
oc
e
e
di
ngs
of
t
he
2nd
C
l
i
ni
c
al
N
at
u
r
al
L
anguage
P
r
oc
e
s
s
i
ng W
or
k
s
hop
,
pp.
72
–
78, M
i
nn
e
a
pol
i
s
, M
i
nn
e
s
ot
a
, U
S
A
,
2019.
[
8]
I
.
B
e
l
t
a
gy,
K
.
L
o,
a
nd
A
.
C
oha
n,
“
S
C
I
B
E
R
T
:
a
pr
e
t
r
a
i
ne
d
l
a
ngua
ge
m
ode
l
f
or
s
c
i
e
nt
i
f
i
c
t
e
xt
,”
i
n
E
M
N
L
P
-
I
J
C
N
L
P
2019
-
2019
C
onf
e
r
e
nc
e
on
E
m
pi
r
i
c
al
M
e
t
hods
i
n
N
at
ur
al
L
anguage
P
r
oc
e
s
s
i
ng
and
9t
h
I
n
t
e
r
nat
i
onal
J
oi
nt
C
onf
e
r
e
nc
e
on
N
at
ur
al
L
anguag
e
P
r
oc
e
s
s
i
ng, P
r
oc
e
e
di
ngs
of
t
he
C
onf
e
r
e
nc
e
, 2019, pp. 3615
–
3620. doi
:
10.18653/
v1/
d19
-
1371.
[
9]
Y
.
L
i
u
e
t
al
.
,
“
R
oB
E
R
T
a
:
a
r
obus
t
l
y
opt
i
m
i
z
e
d
B
E
R
T
pr
e
t
r
a
i
ni
ng
a
ppr
oa
c
h
,”
i
n
I
C
L
R
2020
C
onf
e
r
e
nc
e
P
r
ogr
am
C
hai
r
s
,
pp. 1
–
15
, 2019
.
[
10]
K
.
K
.
B
r
e
s
s
e
m
e
t
al
.
,
“
m
e
dB
E
R
T
.de
:
a
c
om
pr
e
he
ns
i
ve
G
e
r
m
a
n
B
E
R
T
m
od
e
l
f
or
t
he
m
e
di
c
a
l
dom
a
i
n,”
E
x
pe
r
t
Sy
s
t
e
m
s
w
i
t
h
A
ppl
i
c
at
i
ons
, vol
. 237, M
a
r
. 2024, doi
:
10.1016/
j
.e
s
w
a
.2023.121598.
[
11]
S
.
W
a
da
,
T
.
T
a
ke
d
a
,
S
.
M
a
na
be
,
S
.
K
oni
s
hi
,
J
.
K
a
m
oha
r
a
,
a
nd
Y
.
M
a
t
s
um
ur
a
,
“
P
r
e
-
t
r
a
i
ni
ng
t
e
c
hni
que
t
o
l
oc
a
l
i
z
e
m
e
di
c
a
l
B
E
R
T
a
nd e
nha
nc
e
bi
om
e
di
c
a
l
B
E
R
T
,”
a
r
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
,
pp. 1
-
11,
2020
.
[
12]
A
. V
a
s
w
a
ni
e
t
al
.
, “
A
t
t
e
nt
i
on i
s
a
l
l
you ne
e
d,”
i
n
A
dv
anc
e
s
i
n N
e
ur
al
I
nf
or
m
at
i
on P
r
oc
e
s
s
i
ng Sy
s
t
e
m
s
, pp. 5999
–
6009
, 2017
.
[
13]
T
. M
i
kol
ov, M
. K
a
r
a
f
i
á
t
, L
.
B
ur
ge
t
, C
. J
a
n,
a
nd S
. K
huda
npur
,
“
R
e
c
ur
r
e
nt
ne
ur
a
l
ne
t
w
or
k ba
s
e
d
l
a
ngua
ge
m
ode
l
,”
P
r
oc
e
e
di
ngs
of
t
he
11t
h
A
nnual
C
onf
e
r
e
nc
e
of
t
he
I
nt
e
r
nat
i
onal
Spe
e
c
h
C
om
m
uni
c
at
i
on
A
s
s
oc
i
at
i
on,
I
nt
e
r
s
pe
e
c
h
2010
,
pp.
1045
–
1048,
2010,
doi
:
10.21437/
i
nt
e
r
s
pe
e
c
h.2010
-
343.
[
14]
S
.
H
oc
hr
e
i
t
e
r
a
nd
J
.
S
c
hm
i
dhube
r
,
“
L
ong
s
hor
t
-
t
e
r
m
m
e
m
o
r
y,”
N
e
ur
al
C
o
m
put
at
i
on
,
vol
.
9,
no.
8,
pp.
1735
–
1780,
1997,
doi
:
10.1162/
ne
c
o.1997.9.8.1735.
Evaluation Warning : The document was created with Spire.PDF for Python.