I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
4
,
No
.
3
,
J
u
n
e
20
2
4
,
p
p
.
3434
~
3
4
4
2
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
4
i
3
.
pp
3
4
3
4
-
3
4
4
2
3434
J
o
ur
na
l
ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
A multimo
da
l ma
chine learni
ng
appro
a
ch t
o
genera
t
e news
a
rticles fro
m g
eo
-
tag
g
ed
ima
g
es
Abha
y
G
o
t
ma
re
1
,
G
a
nd
ha
rv
a
T
hite
2
,
L
a
x
m
i Bewo
o
r
2
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
Te
c
h
n
o
l
o
g
y
,
V
i
s
h
w
a
k
a
r
m
a
I
n
st
i
t
u
t
e
o
f
I
n
f
o
r
ma
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
P
u
n
e
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
V
i
sh
w
a
k
a
r
ma
I
n
st
i
t
u
t
e
o
f
I
n
f
o
r
ma
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
P
u
n
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
No
v
3
0
,
2
0
2
3
R
ev
is
ed
Ma
r
7
,
2
0
2
4
Acc
ep
ted
Ma
r
9
,
2
0
2
4
Clas
sic
a
l
m
a
c
h
in
e
lea
rn
in
g
a
lg
o
r
it
h
m
s
t
y
p
ica
ll
y
o
p
e
ra
te
o
n
u
n
im
o
d
a
l
d
a
ta
a
n
d
h
e
n
c
e
it
c
a
n
a
n
a
l
y
z
e
a
n
d
m
a
k
e
p
re
d
icti
o
n
s
b
a
se
d
o
n
d
a
ta
fro
m
a
sin
g
le
so
u
rc
e
(m
o
d
a
li
t
y
).
Wh
e
re
a
s
m
u
lt
imo
d
a
l
m
a
c
h
in
e
lea
rn
i
n
g
a
lg
o
r
it
h
m
,
lea
rn
s
fro
m
in
fo
rm
a
ti
o
n
a
c
ro
ss
m
u
lt
ip
le m
o
d
a
li
ti
e
s,
su
c
h
a
s tex
t,
ima
g
e
s,
a
u
d
i
o
,
a
n
d
se
n
so
r
d
a
ta.
T
h
e
p
a
p
e
r
lev
e
ra
g
e
s
th
e
f
u
n
c
ti
o
n
a
li
ti
e
s
o
f
m
u
l
ti
m
o
d
a
l
m
a
c
h
in
e
lea
rn
in
g
(M
L)
a
p
p
li
c
a
ti
o
n
f
o
r
g
e
n
e
ra
ti
n
g
tex
t
fr
o
m
ima
g
e
s.
Th
e
p
ro
p
o
se
d
wo
rk
p
re
se
n
ts
a
n
in
n
o
v
a
ti
v
e
m
u
lt
imo
d
a
l
a
l
g
o
rit
h
m
th
a
t
a
u
to
m
a
tes
th
e
c
re
a
ti
o
n
o
f
n
e
ws
a
rti
c
les
fro
m
g
e
o
-
tag
g
e
d
ima
g
e
s
b
y
lev
e
ra
g
i
n
g
c
u
tt
in
g
-
e
d
g
e
d
e
v
e
lo
p
m
e
n
ts
i
n
m
a
c
h
in
e
lea
rn
i
n
g
,
ima
g
e
c
a
p
ti
o
n
i
n
g
,
a
n
d
a
d
v
a
n
c
e
d
tex
t
g
e
n
e
ra
ti
o
n
tec
h
n
o
lo
g
ies
.
Emp
l
o
y
in
g
a
m
u
lt
im
o
d
a
l
a
p
p
ro
a
c
h
t
h
a
t
in
teg
ra
tes
m
a
c
h
in
e
lea
rn
in
g
a
n
d
tran
sfo
rm
e
r
a
lg
o
rit
h
m
s,
su
c
h
a
s
v
isu
a
l
g
e
o
m
e
try
g
r
o
u
p
n
e
two
rk
1
6
(
VG
G
Ne
t1
6
)
,
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(CNN
)
a
n
d
a
lo
n
g
sh
o
rt
-
term
m
e
m
o
ry
(LS
TM
)
b
a
se
d
sy
ste
m
,
th
e
a
lg
o
rit
h
m
in
it
iate
s
b
y
e
x
trac
ti
n
g
th
e
l
o
c
a
ti
o
n
fr
o
m
e
x
c
h
a
n
g
e
a
b
le
ima
g
e
fil
e
f
o
rm
a
t
(E
x
if)
d
a
ta
fro
m
th
e
ima
g
e
.
Th
e
fe
a
tu
re
s
a
re
e
x
t
ra
c
ted
fro
m
t
h
e
ima
g
e
a
n
d
c
o
rr
e
sp
o
n
d
in
g
n
e
ws
h
e
a
d
li
n
e
is
g
e
n
e
ra
ted
.
T
h
e
h
e
a
d
li
n
e
s
a
re
u
se
d
fo
r
g
e
n
e
ra
ti
n
g
a
c
o
m
p
re
h
e
n
siv
e
a
rti
c
le
wit
h
c
o
n
tem
p
o
ra
ry
larg
e
lan
g
u
a
g
e
m
o
d
e
l
(LL
M
).
F
u
rth
e
r,
th
e
a
lg
o
rit
h
m
g
e
n
e
ra
tes
th
e
n
e
ws
a
rti
c
le
b
ig
-
sc
ien
c
e
larg
e
o
p
e
n
-
sc
ien
c
e
o
p
e
n
-
a
c
c
e
ss
m
u
lt
il
in
g
u
a
l
lan
g
u
a
g
e
m
o
d
e
l
(BLOOM).
Th
e
a
lg
o
rit
h
m
wa
s
tes
ted
o
n
re
a
l
ti
m
e
p
h
o
to
g
ra
p
h
s
a
s
we
ll
a
s
ima
g
e
s
fro
m
th
e
i
n
tern
e
t
.
In
b
o
t
h
t
h
e
c
a
se
s
th
e
n
e
ws
a
rti
c
les
g
e
n
e
ra
ted
we
re
v
a
li
d
a
ted
wi
th
R
OU
G
E
a
n
d
BULE
sc
o
re
.
Th
e
p
ro
p
o
se
d
w
o
rk
is
fo
u
n
d
t
o
b
e
su
c
c
e
ss
fu
l
a
tt
e
m
p
t
in
jo
u
r
n
a
li
sm
field
.
K
ey
w
o
r
d
s
:
B
L
OOM
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
J
o
u
r
n
alis
m
L
ar
g
e
lan
g
u
ag
e
m
o
d
el
L
o
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
Mu
ltimo
d
al
m
a
c
h
in
e
lea
rn
i
n
g
VGGN
et1
6
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
:
L
ax
m
i Bewo
o
r
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
E
n
g
i
n
ee
r
in
g
,
Vis
h
wak
ar
m
a
I
n
s
titu
t
e
o
f
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
Pu
n
e,
I
n
d
ia
E
m
ail: la
x
m
i.b
ewo
o
r
@
v
iit.ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
m
an
n
er
th
at
n
ews
is
d
is
tr
ib
u
ted
an
d
co
n
s
u
m
ed
h
as
s
ig
n
if
ican
tly
ch
an
g
ed
with
th
e
ad
v
en
t
o
f
th
e
d
ig
ital
ag
e.
I
m
a
g
es
h
av
e
b
ec
o
m
e
a
cr
u
cial
co
m
p
o
n
en
t
o
f
th
is
n
ew
en
v
ir
o
n
m
en
t
b
ec
a
u
s
e
o
f
th
eir
i
n
n
ate
ca
p
ac
ity
to
co
n
cisely
ex
p
r
ess
co
m
p
lex
s
to
r
y
lin
es.
T
h
ey
f
r
eq
u
en
tly
co
n
tain
im
p
o
r
ta
n
t
in
f
o
r
m
atio
n
th
at
s
u
p
p
lem
en
ts
tex
tu
al
co
n
ten
t
a
n
d
o
cc
asio
n
ally
ev
en
r
ep
lace
s
it.
Ho
wev
er
,
cr
ea
tin
g
n
ews
s
to
r
ies
m
an
u
ally
f
r
o
m
p
h
o
to
s
is
a
tim
e
-
co
n
s
u
m
in
g
task
th
at
r
eq
u
ir
es
a
lo
t
o
f
lab
o
r
an
d
k
n
o
wled
g
e.
T
r
ad
itio
n
a
l
m
ac
h
in
e
lear
n
in
g
(
ML
)
alg
o
r
ith
m
s
h
a
v
e
lim
itatio
n
s
with
th
e
u
n
if
o
r
m
d
ata
f
o
r
m
at,
b
u
t
in
teg
r
atin
g
m
o
r
e
th
a
n
o
n
e
t
y
p
e
o
f
d
ata
is
th
e
n
ee
d
o
f
th
e
h
o
u
r
.
Mu
ltimo
d
al
f
u
n
ctio
n
alities
[
1
]
,
[
2
]
h
a
v
e
r
ec
en
tly
g
ar
n
e
r
ed
th
e
atten
ti
o
n
o
f
r
esear
c
h
er
s
to
o
v
er
co
m
e
th
is
lim
itatio
n
.
Mu
ltimo
d
al
ap
p
licatio
n
s
h
a
v
e
p
r
o
v
en
to
b
e
e
f
f
ec
tiv
e
f
o
r
h
y
b
r
id
izi
n
g
th
e
m
o
d
els
[
3
]
,
[
4
]
f
o
r
tex
t
an
d
im
ag
e
d
ata
ty
p
es.
T
h
e
p
r
o
p
o
s
ed
wo
r
k
also
r
ep
r
esen
ts
an
en
d
ea
v
o
r
to
d
ev
elo
p
a
m
u
ltimo
d
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
mu
ltimo
d
a
l m
a
ch
in
e
lea
r
n
i
n
g
a
p
p
r
o
a
ch
to
g
e
n
era
te
n
ews
a
r
ticles fr
o
m
…
(
A
b
h
a
y
Go
tma
r
e
)
3435
ap
p
licatio
n
an
d
,
in
tu
r
n
,
s
tep
in
to
th
e
wo
r
ld
o
f
g
en
er
ativ
e
ar
tific
ial
in
tellig
en
ce
(
AI
)
[
5
]
,
[
6
]
.
T
h
er
e
f
o
r
e,
th
e
cr
ea
tio
n
o
f
au
to
m
ate
d
s
y
s
tem
s
ca
p
ab
le
o
f
ca
r
r
y
in
g
o
u
t th
is
ac
tiv
ity
co
u
ld
co
m
p
letely
tr
an
s
f
o
r
m
th
e
jo
u
r
n
alis
m
in
d
u
s
tr
y
b
y
i
n
cr
ea
s
in
g
t
h
e
p
r
o
d
u
ctiv
ity
an
d
s
ca
lab
ilit
y
o
f
n
ews
o
u
tp
u
t.
T
h
e
d
e
v
elo
p
m
en
t
o
f
ar
tific
ial
in
tellig
en
ce
an
d
m
ac
h
in
e
lea
r
n
in
g
tech
n
o
lo
g
ies
[
7
]
,
[
8
]
i
n
r
ec
en
t
y
ea
r
s
h
as
o
p
en
ed
s
ig
n
if
ican
t
o
p
p
o
r
tu
n
ities
f
o
r
a
u
to
m
atin
g
d
if
f
er
e
n
t
p
a
r
ts
o
f
c
o
n
ten
t
g
en
er
atio
n
.
Par
tic
u
lar
ly
im
p
r
ess
iv
e
im
p
r
o
v
e
m
e
n
ts
h
av
e
b
ee
n
m
a
d
e
in
tex
t
g
en
er
atio
n
m
o
d
els
an
d
im
ag
e
ca
p
tio
n
i
n
g
m
eth
o
d
s
.
T
r
an
s
f
o
r
m
er
m
o
d
els
h
a
v
e
b
ec
o
m
e
a
p
o
ten
t
t
o
o
l
in
th
e
f
ield
o
f
n
atu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
,
as
in
tr
o
d
u
ce
d
b
y
Vaswan
i
et
a
l.
[
9
]
.
T
h
ese
m
o
d
els
h
av
e
s
h
o
wn
im
p
r
o
v
e
d
ab
ilit
y
in
co
m
p
r
eh
en
d
in
g
th
e
c
o
n
tex
t
a
n
d
p
r
o
d
u
cin
g
co
h
er
en
t
tex
t
th
a
n
k
s
to
th
eir
s
elf
-
atten
tio
n
m
ec
h
an
is
m
.
L
ater
m
o
d
els h
av
e
in
cr
ea
s
ed
th
eir
u
s
e
o
f
tr
an
s
f
o
r
m
er
s
,
d
em
o
n
s
tr
atin
g
th
e
v
al
u
e
o
f
p
r
e
-
tr
ain
in
g
o
n
h
u
g
e
tex
t
co
r
p
o
r
a
f
o
r
a
v
ar
iet
y
o
f
d
o
wn
s
tr
ea
m
task
s
with
m
o
d
els
lik
e
b
id
ir
ec
tio
n
al
en
c
o
d
er
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
a
n
s
f
o
r
m
er
s
(
B
E
R
T
)
[
1
0
]
–
[
1
2
]
a
n
d
b
ig
-
s
cien
ce
lar
g
e
o
p
en
-
s
cien
ce
o
p
en
-
ac
ce
s
s
m
u
ltil
in
g
u
al
lan
g
u
a
g
e
m
o
d
el
(
B
L
OOM
)
[
1
3
]
.
At
th
e
s
am
e
tim
e,
th
er
e
h
as
b
ee
n
s
u
b
s
tan
tial
d
ev
el
o
p
m
en
t
in
th
e
f
ield
o
f
im
ag
e
ca
p
tio
n
in
g
.
T
h
e
m
o
s
t
co
m
m
o
n
m
o
d
els
[
1
4
]
–
[
1
6
]
a
r
e
t
h
o
s
e
th
at
co
m
b
in
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
[
1
7
]
–
[
1
9
]
f
o
r
v
is
u
al
f
ea
tu
r
e
ex
tr
ac
tio
n
with
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
R
NNs)
o
r
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
n
etwo
r
k
s
f
o
r
ca
p
tio
n
p
r
o
d
u
cti
o
n
.
T
h
e
s
e
alg
o
r
ith
m
s
h
av
e
p
r
o
v
en
to
b
e
ex
ce
llen
t
at
p
r
o
d
u
cin
g
p
er
tin
e
n
t
an
d
p
r
ec
is
e
im
ag
e
ca
p
tio
n
s
.
Desp
ite
t
h
ese
d
ev
elo
p
m
en
ts
,
cr
ea
ti
n
g
n
ews
p
iece
s
f
r
o
m
p
h
o
to
g
r
ap
h
s
s
till
p
r
esen
ts
s
p
e
cif
ic
d
if
f
icu
lties
.
New
s
p
iece
s
ar
e
b
y
n
at
u
r
e
c
o
m
p
lex
;
th
ey
f
r
eq
u
en
tly
ca
ll
f
o
r
a
th
o
r
o
u
g
h
c
o
m
p
r
eh
en
s
io
n
o
f
th
e
co
n
tex
t,
th
e
a
p
titu
d
e
to
r
ec
o
g
n
ize
im
p
o
r
tan
t
e
v
en
ts
,
an
d
th
e
ab
ilit
y
to
wr
ite
in
a
jo
u
r
n
alis
tic
s
ty
le
th
at
ap
p
ea
ls
to
a
w
id
e
au
d
ien
ce
.
T
h
is
d
u
ty
en
tails
m
o
r
e
th
an
ju
s
t
co
m
in
g
u
p
with
a
ca
p
tio
n
f
o
r
an
im
ag
e;
it
also
en
tails
tu
r
n
in
g
th
at
d
escr
ip
tio
n
in
to
a
len
g
th
y
n
ews
r
ep
o
r
t
th
at
o
f
f
er
s
in
-
d
ep
th
d
etails
ab
o
u
t
th
e
p
o
r
tr
ay
e
d
ev
e
n
t.
I
n
t
h
is
r
esear
ch
,
we
ar
e
aim
in
g
to
b
r
id
g
e
th
is
g
a
p
b
y
d
ev
elo
p
i
n
g
a
n
o
v
el
m
o
d
el
f
o
r
g
en
er
atin
g
n
ews
ar
ticles
f
r
o
m
im
ag
es.
Dr
awin
g
u
p
o
n
th
e
te
ch
n
iq
u
es
an
d
i
n
s
ig
h
ts
f
r
o
m
th
e
liter
atu
r
e
in
b
o
th
tex
t
g
en
er
atio
n
an
d
im
ag
e
ca
p
tio
n
in
g
,
we
will
ex
p
l
o
r
e
th
e
p
o
ten
tial
o
f
in
teg
r
atin
g
th
ese
te
ch
n
o
lo
g
ies
f
o
r
o
u
r
p
u
r
p
o
s
e.
Ou
r
g
o
al
is
to
cr
ea
te
a
s
y
s
tem
th
at
ca
n
au
to
m
ati
ca
lly
g
en
er
ate
h
ig
h
-
q
u
ality
,
c
o
n
tex
tu
ally
ac
cu
r
ate
n
ews
ar
ticles
f
r
o
m
im
ag
es,
th
er
eb
y
r
ev
o
lu
tio
n
izin
g
th
e
p
r
o
c
ess
o
f
n
ews
p
r
o
d
u
ctio
n
an
d
d
i
s
s
em
in
atio
n
in
th
e
d
ig
ital
ag
e.
T
h
is
r
esear
ch
h
o
ld
s
th
e
p
o
ten
tial
t
o
co
n
tr
ib
u
te
s
ig
n
if
ican
tly
t
o
th
e
f
ield
s
o
f
au
t
o
m
ated
jo
u
r
n
alis
m
an
d
AI
-
d
r
iv
en
c
o
n
ten
t
g
en
er
at
io
n
[
2
0
]
.
2.
RE
L
AT
E
D
WO
RK
T
h
e
r
ec
en
t
er
a
h
as
witn
ess
ed
a
s
ig
n
if
ican
t
s
h
if
t
f
r
o
m
la
n
g
u
a
g
e
u
n
d
e
r
s
tan
d
in
g
t
o
lan
g
u
ag
e
g
en
er
atio
n
.
I
q
b
al
an
d
Qu
r
esh
i
[
2
0
]
p
r
ese
n
ted
a
d
etailed
s
u
r
v
ey
o
n
tex
t
g
en
er
atio
n
m
o
d
els.
Ad
d
itio
n
ally
,
Am
in
-
Neja
d
et
a
l.
[
2
1
]
u
tili
ze
d
tr
an
s
f
o
r
m
e
r
s
s
u
ch
as
B
E
R
T
an
d
b
id
ir
ec
tio
n
a
l
en
co
d
e
r
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
er
s
(
B
io
B
E
R
T
)
o
n
th
e
m
ed
ical
in
f
o
r
m
atio
n
m
ar
t
f
o
r
in
ten
s
i
v
e
ca
r
e
-
I
I
I
(
MI
MI
C
-
I
I
I
)
d
ataset.
T
h
eir
f
in
d
in
g
s
s
u
g
g
est
th
at
s
y
n
th
etic
d
ata
ca
n
y
ield
b
etter
r
esu
lts
th
an
p
o
o
r
er
q
u
ality
o
r
ig
in
al
d
ata.
Similar
ly
,
Ab
d
elwa
h
ab
an
d
E
lm
a
g
h
r
a
b
y
[
2
2
]
c
o
m
p
a
r
ed
L
STM
,
R
NN,
g
ated
r
ec
u
r
r
en
t
u
n
it
(
GR
U)
,
an
d
Ma
r
k
o
v
ch
ain
m
o
d
els,
co
n
clu
d
in
g
th
at
d
ee
p
lear
n
in
g
m
o
d
els
h
an
d
le
in
c
o
r
r
ec
tly
lab
eled
r
ev
iews
b
etter
th
an
Ma
r
k
o
v
ch
ai
n
-
b
ased
tex
t
g
en
er
ato
r
s
.
A
s
h
if
t
to
war
d
s
s
tr
u
ctu
r
in
g
d
ata
was
o
b
s
er
v
e
d
with
Ma
r
ch
e
g
g
ian
i
a
n
d
Per
ez
-
B
eltr
ac
h
in
i
[
2
3
]
u
s
in
g
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
s
o
n
th
e
web
n
atu
r
al
lan
g
u
ag
e
g
en
er
ato
r
(
W
eb
NL
G)
an
d
SR
1
1
Dee
p
d
atasets
,
d
em
o
n
s
tr
atin
g
th
e
b
e
n
ef
its
o
f
e
n
co
d
i
n
g
s
tr
u
ctu
r
al
i
n
f
o
r
m
atio
n
with
s
g
r
a
p
h
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
s
(
GC
N)
.
Vaswan
i
et
a
l.
[
9
]
in
tr
o
d
u
ce
d
t
h
e
t
r
an
s
f
o
r
m
er
ar
c
h
itectu
r
e
an
d
s
elf
-
atten
tio
n
m
ec
h
an
is
m
s
,
elim
in
atin
g
th
e
n
ee
d
f
o
r
r
ec
u
r
r
e
n
t
lay
e
r
s
.
T
h
e
f
ield
o
f
lar
g
e
lan
g
u
ag
e
m
o
d
el
(
L
L
M)
h
as
also
s
ee
n
r
e
m
ar
k
a
b
le
d
ev
elo
p
m
e
n
ts
.
Scao
et
a
l.
[
1
3
]
in
tr
o
d
u
ce
d
B
L
OOM
,
a
d
ec
o
d
er
-
o
n
ly
tr
a
n
s
f
o
r
m
e
r
lan
g
u
a
g
e
m
o
d
el
th
at
s
h
o
wed
co
m
p
etitiv
e
p
e
r
f
o
r
m
an
ce
a
f
ter
m
u
ltit
ask
f
in
e
-
t
u
n
in
g
.
Hu
et
a
l.
[
2
4
]
p
r
o
p
o
s
ed
l
o
w
-
r
an
k
ad
ap
tatio
n
(L
o
R
A)
,
a
lo
w
-
r
an
k
ad
a
p
tatio
n
o
f
lar
g
e
l
an
g
u
ag
e
m
o
d
els,
wh
ich
g
r
ea
tly
r
ed
u
ce
d
th
e
n
u
m
b
er
o
f
tr
ain
ab
le
p
ar
am
eter
s
f
o
r
d
o
wn
s
tr
ea
m
task
s
.
C
h
ai
et
a
l.
[
2
5
]
p
r
o
p
o
s
ed
a
m
eth
o
d
o
lo
g
y
th
at
s
ig
n
i
f
ican
tly
r
ed
u
ce
d
v
i
d
eo
r
a
n
d
o
m
ac
ce
s
s
m
em
o
r
y
(
VR
AM
)
r
eq
u
ir
em
e
n
ts
,
en
ab
lin
g
f
i
n
e
-
tu
n
i
n
g
o
f
a
7
B
L
L
M
o
n
a
6
GB
co
n
s
u
m
er
-
g
r
a
d
e
g
r
a
p
h
ical
p
r
o
ce
s
s
in
g
u
n
it
(
GPU)
.
T
h
e
a
p
p
licatio
n
o
f
th
ese
m
o
d
els
in
im
ag
e
ca
p
tio
n
in
g
h
as
also
b
ee
n
ex
p
lo
r
e
d
.
Sh
ar
m
a
et
a
l.
[
1
4
]
s
u
g
g
ested
th
at
m
u
lt
im
o
d
al
im
ag
e
ca
p
tio
n
in
g
ca
n
b
e
ef
f
ec
tiv
el
y
d
o
n
e
u
s
in
g
L
STM
,
as
ev
alu
ated
b
y
b
ilin
g
u
al
e
v
alu
atio
n
u
n
d
er
s
tu
d
y
(
B
L
E
U)
s
co
r
es.
Ho
s
s
ain
et
a
l.
[
1
5
]
illu
s
tr
ated
a
n
ar
c
h
itectu
r
e
f
o
r
im
a
g
e
ca
p
tio
n
in
g
u
s
in
g
C
NN
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
Vin
y
als
et
a
l.
[
1
6
]
em
p
h
asized
th
e
u
s
e
o
f
L
STM
n
etwo
r
k
s
with
s
o
f
t
atten
tio
n
m
ec
h
an
is
m
s
.
I
n
co
n
clu
s
io
n
,
th
e
f
iel
d
o
f
tex
t
g
en
er
atio
n
m
o
d
els
an
d
im
ag
e
ca
p
tio
n
in
g
tech
n
iq
u
es
h
as
s
ee
n
s
ig
n
if
ica
n
t
ad
v
a
n
ce
m
en
ts
an
d
a
p
p
licat
io
n
s
.
Ho
wev
er
,
th
er
e
is
s
till
m
u
ch
t
o
ex
p
lo
r
e
an
d
u
n
d
er
s
tan
d
i
n
th
ese
ar
ea
s
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
D
I
n
th
is
s
ec
tio
n
,
th
e
u
s
e
o
f
m
u
ltimo
d
al
al
g
o
r
ith
m
s
is
d
is
cu
s
s
ed
.
T
h
e
m
u
ltimo
d
al
ML
in
teg
r
ates
in
f
o
r
m
atio
n
f
r
o
m
v
ar
io
u
s
m
o
d
alities
lik
e
tex
t,
im
ag
e,
an
d
v
id
eo
.
T
h
e
p
r
o
p
o
s
ed
wo
r
k
i
n
p
u
ts
im
ag
es
a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
4
,
No
.
3
,
J
u
n
e
2
0
2
4
:
3
4
3
4
-
3
4
4
2
3436
g
en
er
ates
r
elev
an
t
tex
t
d
escr
ib
in
g
th
at
im
a
g
e.
T
h
e
en
tire
wo
r
k
f
lo
w
o
f
p
r
o
p
o
s
ed
m
u
ltimo
d
al
ML
alg
o
r
ith
m
s
is
d
ep
icted
in
Fig
u
r
e
1
wh
ic
h
in
c
lu
d
es f
o
llo
win
g
s
tep
s
:
a.
Data
co
llectio
n
an
d
p
r
ep
a
r
atio
n
:
th
e
d
ataset
p
r
im
ar
ily
co
m
p
r
is
es
im
ag
es,
en
co
m
p
ass
in
g
b
o
th
r
ea
l
-
tim
e
im
ag
es
ca
p
tu
r
e
d
with
m
o
b
i
le
ca
m
er
as
an
d
J
PEG
f
iles
o
b
tain
e
d
o
n
lin
e.
T
h
ese
i
m
ag
es
u
n
d
er
g
o
p
r
ep
r
o
ce
s
s
in
g
to
a
f
o
r
m
at
s
u
itab
le
f
o
r
s
u
b
s
eq
u
en
t a
n
aly
s
is
.
b.
C
h
o
ice
o
f
d
ee
p
lear
n
in
g
(
DL
)
m
o
d
els
:
th
is
s
tep
in
v
o
lv
es
s
e
le
ctin
g
th
e
ap
p
r
o
p
r
iate
DL
m
o
d
els
f
o
r
tr
ain
in
g
.
I
n
th
e
c
u
r
r
en
t
wo
r
k
,
th
e
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
-
16
(
VG
G
-
16
)
,
a
C
NN
-
b
ased
m
o
d
el,
is
em
p
lo
y
ed
f
o
r
im
ag
e
p
r
o
ce
s
s
in
g
,
wh
ile
L
STM
an
d
T
r
an
s
f
o
r
m
er
-
b
ase
d
m
o
d
els
h
an
d
le
tex
tu
al
d
ata
p
r
o
ce
s
s
in
g
.
Ad
d
itio
n
ally
,
th
e
lar
g
e
lan
g
u
a
g
e
m
o
d
el
B
L
OOM
-
1
b
7
p
r
e
-
tr
ain
ed
m
o
d
el
is
u
tili
ze
d
f
o
r
te
x
t e
n
h
an
ce
m
e
n
t.
c.
C
h
o
ice
o
f
DL
m
o
d
els:
th
is
s
tep
in
v
o
l
v
es
s
elec
tin
g
th
e
ap
p
r
o
p
r
iate
DL
m
o
d
el
f
o
r
tr
ain
in
g
.
I
n
th
e
cu
r
r
e
n
t
wo
r
k
,
VGG
-
1
6
,
a
C
NN
-
b
ased
m
o
d
el,
is
u
tili
ze
d
f
o
r
im
ag
e
p
r
o
ce
s
s
in
g
,
wh
ile
L
STM
a
n
d
t
r
an
s
f
o
r
m
er
-
b
ased
m
o
d
els ar
e
em
p
lo
y
e
d
f
o
r
tex
tu
al
d
ata
p
r
o
ce
s
s
in
g
.
d.
Mo
d
el
h
y
b
r
id
izatio
n
:
th
is
s
tep
f
ac
ilit
ates
lev
er
ag
in
g
th
e
a
d
v
an
tag
es
o
f
m
u
ltip
le
m
o
d
alities
.
T
h
e
m
o
d
el
u
n
d
er
g
o
es f
u
r
t
h
er
tr
ain
i
n
g
an
d
f
in
e
-
tu
n
i
n
g
b
y
in
teg
r
atin
g
f
ea
t
u
r
es f
r
o
m
b
o
th
i
m
ag
es a
n
d
te
x
t.
e.
I
n
ter
p
r
etatio
n
o
f
th
e
r
esu
lts
:
t
h
e
r
esu
lts
ar
e
as
s
es
s
ed
u
s
in
g
w
ell
-
estab
lis
h
ed
p
er
f
o
r
m
an
ce
m
etr
ics:
ac
cu
r
ac
y
an
d
lo
s
s
r
ate.
T
h
e
ac
cu
r
ac
y
ass
e
s
s
m
en
t
in
clu
d
es
B
L
E
U
an
d
r
ec
all
-
o
r
ien
ted
u
n
d
er
s
tu
d
y
f
o
r
g
is
tin
g
ev
alu
atio
n
(
R
OUGE
)
s
co
r
es
[
2
6
]
.
Alg
o
r
ith
m
1
p
r
o
v
id
es
d
etailed
ex
ec
u
tio
n
s
tep
s
f
o
r
co
n
v
er
tin
g
in
p
u
t im
ag
e
to
tex
t.
Fig
u
r
e
1
.
Pro
p
o
s
ed
ar
c
h
itectu
r
e
d
iag
r
am
Alg
o
r
ith
m
1
.
Pro
p
o
s
ed
atten
tio
n
-
b
ased
h
y
b
r
id
C
NN
-
L
STM
m
o
d
el
Step
1
.
Upload Image: The
user uploads a geo
-
tagged image.
Step
2
.
Extract
Geo
-
ta
gg
e
d
Lo
ca
ti
on
s:
Th
e
e
x
c
h
a
n
g
e
a
b
l
e
i
ma
g
e
f
i
l
e
f
o
r
ma
t
(
Ex
i
f
)
decryption
algorithm
is
used to extract the geo
-
tagged locations from the uploaded image.
Step
3
.
Feature
Extraction:
The
image
is
then
pr
ocessed
for
feature
extraction
using
the
VGGNet16 CN
N
model which is trained with a customized dataset.
Step
4
.
Tu
ne
Ex
tr
ac
te
d
Fe
at
ur
es
:
Th
e
fe
at
ur
es
ex
tr
ac
te
d
fr
om
th
e
im
ag
e
ar
e
th
en
fi
ne
-
tuned
with the trained weights of the LSTM algorithm.
Step
5
.
Generate
Raw
Head
line:
The
output
of
the
L
STM
is
a
raw
headline
tha
t
describes
the
image in a basic way.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
mu
ltimo
d
a
l m
a
ch
in
e
lea
r
n
i
n
g
a
p
p
r
o
a
ch
to
g
e
n
era
te
n
ews
a
r
ticles fr
o
m
…
(
A
b
h
a
y
Go
tma
r
e
)
3437
Step
6
.
Convert
Raw
Headline:
This
raw
headline
is
passed
through
another
LSTM
model
to
convert
the
sentence
into
a
meaningful
headline,
which
will
serve
as
the
actual
headline for the image.
Step
7
.
Pa
ss
He
ad
li
ne
Th
ro
ug
h
fi
ne
tu
ne
d
LL
M:
Th
e
he
ad
li
ne
an
d
th
e
pr
ev
i
ou
sl
y
ex
tr
ac
te
d
location
is
then
passed
through
a
fine
-
tu
ne
d
LL
M
Bl
oo
m
-
1b
7
mo
de
l.
in
th
i
s
ca
se
,
th
e
Bloom
-
1b
7
mo
de
l.
Th
is
mo
d
el
,
fi
ne
-
tu
ne
d
us
in
g
qu
an
ti
ze
d
lo
w
ra
nk
ad
aptation
(QLoRA),
is designed to generate news articles.
Step
8
.
Generate
Description:
The
Bloom
-
1b
7
mo
de
l
ge
ne
ra
te
s
a
me
an
in
gf
ul
de
sc
ri
pt
io
n
fo
r
the headline.
Step
9
.
Combine
Elements
to
Form
Article:
Final
ly,
this
description,
th
e
headline,
the
location
extracted
from
t
he
Exif
decryption
algori
thm,
and
the
image
are
c
ombined
t
o
form a full news article.
End Procedure.
4.
L
E
A
RNING
M
O
DE
L
S O
F
P
RO
P
O
SE
D
M
E
T
H
O
DO
L
O
G
Y
4
.
1
.
E
x
t
ra
ct
io
n o
f
lo
ca
t
io
n c
o
-
o
r
din
a
t
es
T
h
e
g
eo
tag
g
e
d
im
ag
e
h
as
an
en
cr
y
p
ted
E
x
if
f
ile
th
at
is
u
s
ed
to
s
to
r
e
th
e
lo
ca
tio
n
an
d
a
v
ar
iety
o
f
o
th
er
im
ag
e
d
etails.
T
h
e
E
x
if
f
ile
is
d
ec
r
y
p
ted
u
s
in
g
th
e
Py
th
o
n
lan
g
u
a
g
e
f
u
n
ctio
n
alities
to
ex
tr
ac
t th
e
lo
ca
tio
n
co
o
r
d
in
ates.
T
h
e
lo
ca
tio
n
c
o
o
r
d
in
ates
ar
e
r
e
p
r
esen
ted
in
th
e
f
o
r
m
o
f
d
ir
ec
tio
n
s
(
NSEW)
th
at
ar
e
th
en
co
n
v
er
ted
in
to
r
esp
ec
tiv
e
latit
u
d
e
an
d
lo
n
g
itu
d
e
with
th
e
f
o
l
lo
win
g
f
o
r
m
u
las:
=
+
/
60
+
/
3600
=
+
/
60
+
/
3600
T
h
e
co
n
v
er
ted
latitu
d
e
a
n
d
lo
n
g
itu
d
e
ar
e
u
s
ed
to
ex
t
r
ac
t
th
e
ex
ac
t
lo
ca
tio
n
,
a
n
d
a
r
ea
l
-
ti
m
e
m
ap
is
r
en
d
e
r
ed
with
th
e
h
elp
o
f
f
o
liu
m
,
as sh
o
wn
in
Fig
u
r
e
2
.
Fig
u
r
e
2
.
E
x
ac
t e
x
tr
ac
ted
l
o
ca
t
io
n
o
n
f
o
liu
m
m
ap
with
b
l
u
e
m
ar
k
er
4
.
2
.
VG
G
-
1
6
a
nd
L
ST
M
f
o
r
g
ener
a
t
ing
ca
ptio
n f
r
o
m
i
ma
g
e
Af
ter
ex
tr
ac
tin
g
th
e
lo
ca
tio
n
in
f
o
r
m
atio
n
f
r
o
m
th
e
Ex
if
d
ata
o
f
th
e
im
ag
e,
th
e
im
a
g
e
is
th
en
p
ass
ed
to
a
VGGN
et
-
1
6
m
o
d
el
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
.
T
h
e
ty
p
es
an
d
n
u
m
b
er
o
f
p
a
r
am
eter
s
ar
e
s
h
o
wn
in
T
ab
le
1
.
T
h
e
ex
tr
ac
ted
f
ea
tu
r
es
ar
e
s
u
b
s
eq
u
en
tly
s
to
r
ed
an
d
p
ass
ed
in
to
th
e
lo
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
(
L
STM
)
m
o
d
el
f
o
r
tr
ain
in
g
to
ex
tr
ac
t
ca
p
tio
n
s
f
r
o
m
im
ag
es.
T
h
e
VGGN
et
-
1
6
in
n
er
m
o
d
el
ar
c
h
itectu
r
e,
b
a
s
ed
o
n
th
e
r
eq
u
ir
e
d
p
ar
am
eter
s
an
d
f
ea
tu
r
es,
is
illu
s
tr
ated
in
Fig
u
r
e
3
.
T
h
e
c
o
m
p
o
n
en
ts
o
f
th
e
r
esp
ec
tiv
e
L
STM
m
o
d
el
ar
e
p
r
esen
ted
in
T
a
b
le
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
4
,
No
.
3
,
J
u
n
e
2
0
2
4
:
3
4
3
4
-
3
4
4
2
3438
T
ab
le
1
.
VGGN
et
-
16
p
ar
am
et
er
d
is
tr
ib
u
tio
n
S
r
N
o
P
a
r
a
me
t
e
r
t
y
p
e
N
u
mb
e
r
o
f
p
a
r
a
me
t
e
r
s
1
To
t
a
l
P
a
r
a
m
e
t
e
r
s
1
3
4
2
6
0
5
4
4
2
Tr
a
i
n
a
b
l
e
P
a
r
a
met
e
r
s
1
3
4
2
6
0
5
4
4
3
N
o
n
-
t
r
a
i
n
a
b
l
e
P
a
r
a
me
t
e
r
s
0
Fig
u
r
e
3
.
VGGN
et
-
16
i
n
n
er
a
r
ch
itectu
r
e
T
ab
le
2
.
L
STM
f
o
r
ca
p
tio
n
g
e
n
er
atio
n
attr
ib
u
tes
Ep
o
c
h
s
B
a
t
c
h
S
i
z
e
Lo
ss
B
LEU
-
1
60
3
0
.
2
3
9
9
0
.
2
9
5
4
.
3
.
L
ST
M
f
o
r
hea
dli
ne
g
en
er
a
t
io
n
Up
o
n
th
e
g
e
n
er
atio
n
o
f
an
in
i
tial
ca
p
tio
n
th
r
o
u
g
h
th
e
c
o
m
b
in
ed
ef
f
o
r
ts
o
f
th
e
VGGNet
-
1
6
an
d
th
e
p
r
im
ar
y
L
STM
m
o
d
el,
it
is
th
en
f
ed
in
to
a
s
ec
o
n
d
a
r
y
L
ST
M
n
etwo
r
k
.
T
h
is
s
u
b
s
eq
u
e
n
t
L
STM
is
s
p
ec
if
ically
en
g
in
ee
r
ed
to
d
is
till
th
e
p
r
eli
m
in
ar
y
ca
p
tio
n
in
t
o
a
co
n
tex
t
u
ally
r
elev
an
t
an
d
s
em
an
ticall
y
co
h
er
en
t
h
ea
d
lin
e.
T
h
e
s
ec
o
n
d
L
STM
m
o
d
el
h
as
th
e
f
o
llo
win
g
attr
ib
u
tes
tab
u
lated
in
T
ab
le
3
.
T
h
e
c
o
m
p
ilatio
n
attr
ib
u
tes
o
f
th
e
L
STM
m
o
d
el
ar
e
tab
u
lated
in
T
ab
le
4
.
T
ab
le
3
.
L
STM
f
o
r
h
ea
d
lin
e
g
en
er
atio
n
attr
ib
u
tes
La
y
e
r
(
t
y
p
e
)
O
u
t
p
u
t
s
h
a
p
e
N
o
o
f
p
a
r
a
me
t
e
r
s
Emb
e
d
d
i
n
g
(
N
o
n
e
,
7
6
,
1
0
0
)
5
4
3
0
0
LSTM
(
N
o
n
e
,
1
5
0
)
1
5
0
6
0
0
D
e
n
se
(
N
o
n
e
,
5
4
3
)
8
1
9
9
3
T
ab
le
4
.
L
STM
co
m
p
ilatio
n
at
tr
ib
u
tes
Ep
o
c
h
s
B
a
t
c
h
S
i
z
e
Lo
ss
A
c
c
u
r
a
c
y
2
0
0
1
0
.
0
1
3
0
0
.
9
9
1
6
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
mu
ltimo
d
a
l m
a
ch
in
e
lea
r
n
i
n
g
a
p
p
r
o
a
ch
to
g
e
n
era
te
n
ews
a
r
ticles fr
o
m
…
(
A
b
h
a
y
Go
tma
r
e
)
3439
4
.
4
.
L
L
M
f
o
r
g
ener
a
t
ing
ne
ws f
ro
m
hea
dli
ne
Af
ter
g
en
er
atin
g
th
e
h
ea
d
lin
e
,
it
is
p
as
s
ed
to
th
e
B
lo
o
m
-
1
b
7
p
r
e
-
tr
ain
ed
m
o
d
el,
an
o
p
en
-
s
o
u
r
ce
L
L
M.
T
h
is
m
o
d
el,
with
1
,
7
2
2
,
4
0
8
,
9
6
0
p
a
r
am
eter
s
,
h
as b
ee
n
tr
ain
ed
o
n
4
5
n
atu
r
al
lan
g
u
ag
e
s
an
d
1
.
5
ter
ab
y
tes
o
f
p
r
e
-
p
r
o
ce
s
s
ed
tex
t,
co
n
v
er
ted
in
to
3
5
0
b
illi
o
n
u
n
i
q
u
e
to
k
en
s
.
T
h
e
m
o
d
el
co
m
p
r
is
es
5
1
3
,
8
0
2
,
2
4
0
em
b
ed
d
in
g
p
ar
am
eter
s
,
2
4
lay
er
s
,
1
6
atten
tio
n
h
ea
d
s
,
an
d
2
0
4
8
-
d
im
en
s
io
n
al
h
id
d
en
lay
er
s
.
I
t
is
d
esig
n
ed
to
p
er
f
o
r
m
g
e
n
er
al
-
p
u
r
p
o
s
e
task
s
an
d
ca
n
b
e
f
i
n
e
-
tu
n
e
d
f
o
r
d
o
m
ain
-
s
p
ec
if
ic
task
s
.
T
h
e
f
in
e
-
tu
n
in
g
p
r
o
ce
s
s
was
p
er
f
o
r
m
ed
o
n
a
s
in
g
le
T
esla
T
4
GPU
with
1
6
g
ig
ab
y
tes
o
f
GPU
VR
AM
.
Du
e
to
h
ar
d
war
e
co
n
s
tr
ain
ts
,
we
em
p
lo
y
ed
a
QL
o
R
A
m
eth
o
d
.
T
h
is
m
eth
o
d
in
v
o
lv
ed
r
ed
u
c
in
g
th
e
p
r
ec
is
io
n
o
f
m
o
d
el
p
ar
am
eter
s
to
4
-
b
i
t
in
teg
er
s
,
s
ig
n
if
ican
tly
r
ed
u
cin
g
th
e
am
o
u
n
t
o
f
C
PU
an
d
GPU
m
em
o
r
y
r
eq
u
i
r
ed
,
m
ak
i
n
g
it
f
ea
s
ib
le
to
p
er
f
o
r
m
th
e
ex
p
e
r
im
en
t
o
n
th
e
av
ailab
le
h
ar
d
war
e
.
T
h
e
d
ataset
u
s
ed
f
o
r
f
in
e
-
tu
n
in
g
was
s
elf
-
g
en
e
r
ated
an
d
c
o
n
s
is
ted
o
f
h
ea
d
lin
es a
n
d
th
eir
r
esp
ec
tiv
e
n
ews a
r
ticles.
Desp
ite
its
s
m
all
s
ize,
th
is
d
ataset
s
er
v
ed
as a
p
r
o
to
ty
p
e
f
o
r
t
h
e
ex
p
er
im
en
t.
T
h
e
f
in
e
-
tu
n
in
g
p
r
o
ce
s
s
was
g
u
id
ed
b
y
a
p
r
o
m
p
t
tem
p
late,
wh
ich
was
d
es
ig
n
ed
to
g
en
er
ate
a
n
ews
ar
ticle
b
ased
o
n
a
g
iv
e
n
h
ea
d
lin
e
.
T
h
e
tem
p
late
was
a
s
f
o
llo
ws:
'
B
ased
o
n
th
e
f
o
llo
win
g
h
ea
d
lin
e
a
n
d
lo
ca
tio
n
,
g
en
er
ate
a
n
ews
ar
ticle:
"He
ad
lin
e"
"lo
ca
tio
n
"
"n
ews".
T
h
e
p
ar
am
eter
s
f
o
r
g
en
er
atio
n
wer
e
s
et
with
a
tem
p
er
atu
r
e
o
f
0
.
1
a
n
d
a
m
ax
im
u
m
o
f
5
1
2
n
ew
to
k
en
s
.
T
h
e
ad
ap
ter
ty
p
e
u
s
ed
was
L
o
R
A,
an
d
th
e
q
u
an
tizatio
n
was
s
et
to
4
b
its
.
T
h
ese
p
ar
am
ete
r
s
wer
e
ch
o
s
e
n
to
b
ala
n
ce
th
e
tr
ad
e
-
o
f
f
b
et
wee
n
co
m
p
u
tatio
n
al
ef
f
icien
cy
an
d
th
e
q
u
ality
o
f
t
h
e
g
en
er
ate
d
tex
t.
5.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
th
e
r
ea
lm
o
f
waste
m
a
n
ag
em
en
t,
th
e
in
te
g
r
atio
n
o
f
m
u
ltimo
d
al
ap
p
r
o
ac
h
es
a
d
v
o
ca
tes
a
tr
an
s
f
o
r
m
ativ
e
s
h
if
t
i
n
h
o
w
n
ews
r
elate
d
to
th
is
cr
itical
f
ield
is
g
en
er
ated
a
n
d
d
is
s
em
in
ated
.
T
h
e
p
r
o
p
o
s
ed
wo
r
k
n
o
t
o
n
l
y
wo
r
k
s
o
n
th
e
im
ag
es
av
ailab
le
o
n
lin
e
b
u
t
also
ca
n
co
llect
r
ea
l
tim
e
im
ag
es,
id
en
tifie
s
th
e
lo
ca
tio
n
an
d
r
ele
v
an
t
in
f
o
r
m
a
tio
n
f
r
o
m
t
h
at
p
ictu
r
e
is
g
ettin
g
g
en
e
r
ated
wh
ich
will
b
e
s
u
p
p
lied
to
h
ig
h
e
r
au
th
o
r
ities
f
o
r
f
u
r
th
e
r
ac
tio
n
.
T
h
e
d
etailed
an
al
y
s
is
o
f
r
esu
lts
is
p
r
o
v
id
ed
in
s
u
b
s
eq
u
en
t su
b
s
ec
tio
n
s
.
5
.
1
.
T
he
re
s
ults f
o
r
t
he
ca
pti
o
n g
ener
a
t
ed
f
ro
m
t
he
ima
g
e
T
h
e
m
eth
o
d
o
lo
g
y
f
o
r
th
is
task
in
v
o
lv
e
d
th
e
u
tili
za
tio
n
o
f
a
c
o
m
b
in
atio
n
o
f
VGGN
et
-
1
6
an
d
L
STM
.
T
h
e
co
m
b
in
atio
n
u
n
d
er
wen
t
t
esti
n
g
with
v
ar
io
u
s
h
y
p
er
p
ar
a
m
eter
s
,
an
d
ad
ju
s
tm
en
ts
wer
e
m
ad
e
f
o
r
o
p
tim
al
r
esu
lts
.
W
h
ile
th
e
n
u
m
b
er
o
f
ep
o
ch
s
r
em
ain
ed
c
o
n
s
tan
t
at
6
0
,
we
ex
p
er
im
en
te
d
with
d
if
f
e
r
en
t
b
atch
s
izes
f
o
r
th
e
m
o
d
el,
r
esu
ltin
g
in
th
e
f
i
n
d
in
g
s
tab
u
lated
in
T
a
b
le
5
.
B
ased
o
n
th
e
B
L
E
U
-
1
an
d
B
L
E
U
-
2
s
co
r
es,
th
e
o
p
tim
al
b
atch
s
ize
in
ter
m
s
o
f
ac
cu
r
ac
y
was th
r
ee
,
w
h
er
e
B
L
E
U
-
1
s
co
r
e
=
0
.
2
9
5
an
d
B
L
E
U
-
2
s
co
r
e
=
0
.
T
ab
le
5
.
Hy
p
er
p
a
r
am
eter
tu
n
in
g
r
esu
lts
S
r
.
n
o
Ep
o
c
h
B
a
t
c
h
si
z
e
B
l
e
u
1
B
l
e
u
2
1
60
10
0
.
2
9
0
2
60
20
0
.
3
5
0
3
60
11
0
.
2
9
0
4
60
5
0
.
3
0
0
5
60
15
0
.
3
2
0
.
0
7
6
60
7
0
.
3
1
0
.
0
7
7
60
3
0
.
2
9
0
8
60
4
0
.
2
9
0
9
60
6
0
.
3
2
0
.
0
7
10
60
8
0
.
3
1
0
11
60
9
0
.
3
0
0
.
0
7
5
.
2
.
T
he
re
s
ults f
o
r
t
he
hea
dli
ne
g
ener
a
t
ed
f
ro
m
t
he
ca
pt
io
n
T
h
e
m
eth
o
d
o
lo
g
y
f
o
r
th
is
task
in
v
o
lv
e
d
u
s
in
g
L
STM
to
tr
ain
o
n
th
e
ca
p
tio
n
s
g
en
e
r
ated
f
r
o
m
th
e
im
ag
e,
co
n
v
er
tin
g
th
e
m
in
t
o
s
u
itab
le
h
ea
d
lin
es
f
o
r
n
ews
a
r
ticle
g
en
er
atio
n
.
T
h
e
m
o
d
el'
s
tr
ain
in
g
p
ar
am
eter
s
,
as we
ll a
s
it
s
r
esp
ec
tiv
e
ac
cu
r
ac
y
an
d
lo
s
s
f
ac
to
r
s
,
wer
e
o
b
tain
ed
an
d
tab
u
lated
.
T
ab
le
6
d
e
p
icts
th
e
n
u
m
b
er
o
f
ep
o
ch
s
alo
n
g
with
th
e
co
r
r
esp
o
n
d
in
g
lo
s
s
f
ac
to
r
a
n
d
ac
c
u
r
a
cy
.
C
ateg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
is
co
n
s
id
er
ed
as
th
e
lo
s
s
f
ac
to
r
;
h
e
n
ce
,
2
0
0
ep
o
ch
s
wer
e
c
h
o
s
en
f
o
r
m
o
d
el
c
o
n
s
tr
u
ctio
n
.
T
h
e
ac
cu
r
ac
y
an
d
l
o
s
s
f
ac
to
r
p
er
ep
o
c
h
is
also
p
lo
tted
,
as
s
h
o
wn
in
Fig
u
r
e
s
4
an
d
5
r
esp
ec
tiv
ely
.
T
h
u
s
,
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
d
u
r
in
g
tr
ain
in
g
is
ev
alu
ated
.
5
.
3
.
T
he
re
s
ults f
o
r
t
he
a
rt
ic
le
g
ener
a
t
ed
f
ro
m
t
he
hea
dli
ne
T
h
e
p
r
o
p
o
s
e
d
a
l
g
o
r
i
t
h
m
a
ls
o
m
a
k
e
s
u
s
e
o
f
f
i
n
e
-
t
u
n
i
n
g
B
l
o
o
m
-
1
b
7
.
A
s
s
h
o
w
n
i
n
T
a
b
le
7
,
th
e
f
i
n
d
i
n
g
s
e
n
c
o
m
p
a
s
s
a
n
a
r
r
a
y
o
f
a
c
c
u
r
a
c
y
m
e
t
r
i
c
s
,
i
n
c
l
u
d
i
n
g
d
i
f
f
e
r
e
n
t
v
e
r
s
i
o
n
s
o
f
b
o
t
h
B
L
E
U
a
n
d
R
OU
G
E
s
c
o
r
es
.
T
h
e
s
e
r
e
s
u
l
ts
f
r
o
m
a
l
l
t
h
e
t
h
r
e
e
s
u
b
d
o
m
a
i
n
s
s
e
r
v
e
a
s
t
h
e
c
o
n
c
l
u
s
i
v
e
ev
a
l
u
a
t
i
o
n
o
f
t
h
e
m
o
d
e
l
’
s
p
e
r
f
o
r
m
a
n
c
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
4
,
No
.
3
,
J
u
n
e
2
0
2
4
:
3
4
3
4
-
3
4
4
2
3440
Ou
r
r
esear
ch
m
ar
k
s
th
e
f
ir
s
t
en
d
ea
v
o
r
in
th
e
d
o
m
ain
o
f
g
en
er
atin
g
tex
t
d
ir
ec
tly
f
r
o
m
im
ag
es,
p
io
n
ee
r
in
g
a
n
o
v
el
ap
p
r
o
ac
h
to
b
r
id
g
in
g
t
h
e
g
a
p
b
etwe
en
v
is
u
al
an
d
tex
t
u
al
m
o
d
alities
.
T
h
r
o
u
g
h
ex
ten
s
iv
e
ex
p
er
im
en
tatio
n
an
d
e
v
alu
ati
o
n
,
we
in
tr
o
d
u
ce
a
n
in
n
o
v
ati
v
e
alg
o
r
ith
m
ca
p
ab
le
o
f
au
to
n
o
m
o
u
s
ly
g
en
er
atin
g
d
escr
ip
tiv
e
tex
t
f
o
r
im
ag
es
wi
th
r
em
ar
k
ab
le
ac
c
u
r
ac
y
an
d
c
o
h
er
en
ce
.
T
h
is
g
r
o
u
n
d
b
r
ea
k
in
g
ac
h
iev
em
en
t
n
o
t
o
n
ly
f
ills
a
s
ig
n
if
ican
t
g
ap
i
n
ex
is
tin
g
liter
atu
r
e
b
u
t
also
o
p
en
s
av
e
n
u
es
f
o
r
n
o
v
el
ap
p
licatio
n
s
in
im
ag
e
u
n
d
er
s
tan
d
i
n
g
,
co
n
ten
t
g
e
n
er
a
tio
n
,
an
d
ac
ce
s
s
ib
ilit
y
en
h
an
c
em
en
t.
Ou
r
r
esu
lts
s
h
o
wca
s
e
th
e
tr
an
s
f
o
r
m
ativ
e
p
o
ten
tial
o
f
m
u
ltimo
d
al
lear
n
i
n
g
tec
h
n
iq
u
es
in
u
n
lo
c
k
in
g
n
e
w
ca
p
ab
ilit
ies
f
o
r
u
n
d
er
s
tan
d
i
n
g
a
n
d
in
ter
p
r
eti
n
g
v
is
u
al
co
n
ten
t,
lay
in
g
th
e
f
o
u
n
d
atio
n
f
o
r
f
u
t
u
r
e
ad
v
an
ce
m
e
n
ts
in
th
is
em
er
g
in
g
f
ield
.
T
ab
le
6
.
Hea
d
lin
e
r
esu
lts
S
r
N
o
Ep
o
c
h
s
C
a
t
e
g
o
r
i
c
a
l
C
r
o
ss
-
e
n
t
r
o
p
y
A
c
c
u
r
a
c
y
1
10
4
.
8
4
0
.
1
0
2
20
2
.
6
6
0
.
4
5
3
50
0
.
1
9
0
.
9
8
4
75
0
.
0
6
0
.
9
9
5
1
0
0
0
.
0
3
0
.
9
9
6
1
2
0
0
.
0
2
0
.
9
9
7
2
0
0
0
.
0
1
0
.
9
9
Fig
u
r
e
4
.
Acc
u
r
ac
y
p
er
ep
o
ch
Fig
u
r
e
5
.
L
o
s
s
p
er
ep
o
c
h
T
ab
le
7
.
Fin
e
-
tu
n
ed
b
lo
o
m
m
o
d
el
ac
cu
r
ac
y
p
ar
am
eter
s
S
r
N
o
P
a
r
a
me
t
e
r
S
c
o
r
e
1
B
l
e
u
1
0
.
2
6
2
c
h
a
r
_
e
r
r
o
r
_
r
a
t
e
0
.
7
3
3
l
o
ss
2
.
2
4
4
n
e
x
t
_
t
o
k
e
n
_
p
e
r
p
l
e
x
i
t
y
1
5
2
9
7
1
.
7
6
5
p
e
r
p
l
e
x
i
t
y
2
5
0
8
4
9
.
7
8
6
r
o
u
g
e
1
_
f
mea
s
u
r
e
0
.
5
3
7
r
o
u
g
e
1
_
p
r
e
c
i
s
i
o
n
0
.
4
5
8
r
o
u
g
e
1
_
r
e
c
a
l
l
0
.
6
9
9
r
o
u
g
e
2
_
f
mea
s
u
r
e
0
.
3
5
10
r
o
u
g
e
2
_
p
r
e
c
i
s
i
o
n
0
.
2
8
11
r
o
u
g
e
2
_
r
e
c
a
l
l
0
.
4
6
12
r
o
u
g
e
L
_
f
me
a
s
u
r
e
0
.
4
7
13
r
o
u
g
e
L
_
p
r
e
c
i
si
o
n
0
.
3
9
14
r
o
u
g
e
L
_
r
e
c
a
l
l
0
.
6
1
15
r
o
u
g
e
L
su
m
_
f
me
a
su
r
e
0
.
5
1
16
r
o
u
g
e
L
su
m
_
p
r
e
c
i
si
o
n
0
.
4
2
17
r
o
u
g
e
L
su
m
_
r
e
c
a
l
l
0
.
6
6
18
c
o
m
b
i
n
e
d
_
l
o
ss
2
.
2
4
6.
CO
NCLU
SI
O
N
T
h
is
r
esear
ch
aim
ed
to
d
ev
el
o
p
a
n
o
v
el
alg
o
r
ith
m
f
o
r
g
e
n
er
atin
g
n
ews
ar
ticles
f
r
o
m
g
eo
-
tag
g
e
d
im
ag
es,
lev
er
ag
in
g
ad
v
a
n
ce
m
en
ts
in
m
ac
h
in
e
lear
n
in
g
,
im
ag
e
ca
p
tio
n
in
g
,
an
d
tex
t
g
en
e
r
atio
n
m
o
d
els.
T
h
e
r
esu
lts
o
f
th
e
s
tu
d
y
s
h
o
wed
p
r
o
m
is
in
g
o
u
tc
o
m
es
ac
r
o
s
s
th
e
th
r
ee
s
u
b
d
o
m
ain
s
o
f
th
e
alg
o
r
ith
m
.
T
h
e
u
s
e
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
mu
ltimo
d
a
l m
a
ch
in
e
lea
r
n
i
n
g
a
p
p
r
o
a
ch
to
g
e
n
era
te
n
ews
a
r
ticles fr
o
m
…
(
A
b
h
a
y
Go
tma
r
e
)
3441
VGGNet
-
1
6
an
d
L
STM
f
o
r
i
m
ag
e
ca
p
tio
n
in
g
y
ield
ed
o
p
ti
m
al
r
esu
lts
with
a
b
atch
s
ize
o
f
th
r
ee
,
as
ev
id
en
ce
d
b
y
th
e
B
L
E
U
-
1
an
d
B
L
E
U
-
2
s
co
r
es.
T
h
e
s
u
b
s
eq
u
en
t
c
o
n
v
e
r
s
io
n
o
f
th
e
ca
p
tio
n
in
t
o
a
s
u
itab
le
h
ea
d
lin
e
u
s
in
g
L
STM
d
em
o
n
s
tr
ated
h
ig
h
ac
cu
r
ac
y
,
with
th
e
m
o
d
el
tr
ain
e
d
f
o
r
2
0
0
e
p
o
ch
s
.
Fin
ally
,
th
e
g
en
er
atio
n
o
f
th
e
ar
ticle
f
r
o
m
t
h
e
h
ea
d
lin
e
u
s
i
n
g
a
f
in
e
-
tu
n
e
d
B
lo
o
m
-
1
B
7
L
L
M
r
esu
lted
in
s
atis
f
ac
to
r
y
ac
cu
r
ac
y
m
etr
ics,
in
clu
d
in
g
B
L
E
U
an
d
R
ec
all
-
Or
ien
ted
Un
d
er
s
tu
d
y
f
o
r
R
OUGE
s
co
r
es.
T
h
e
s
tu
d
y
'
s
f
in
d
in
g
s
h
ig
h
lig
h
t
th
e
p
o
ten
tial
o
f
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
in
r
ev
o
lu
tio
n
izin
g
n
ews
p
r
o
d
u
ctio
n
an
d
d
is
s
em
in
atio
n
in
th
e
d
ig
ital
ag
e.
B
y
au
to
m
atin
g
th
e
p
r
o
ce
s
s
o
f
g
en
er
atin
g
n
ews
ar
ticles
f
r
o
m
im
ag
es,
th
e
alg
o
r
ith
m
ca
n
s
ig
n
if
ican
tly
r
ed
u
ce
th
e
m
an
u
al
e
f
f
o
r
t
a
n
d
tim
e
r
eq
u
ir
ed
,
t
h
er
eb
y
im
p
r
o
v
in
g
t
h
e
ef
f
icien
cy
an
d
s
ca
lab
ilit
y
o
f
n
ews
p
r
o
d
u
ctio
n
.
T
h
is
r
esear
ch
r
e
p
r
esen
ts
a
s
i
g
n
if
ican
t
s
tep
t
o
war
d
s
th
e
a
u
to
m
atio
n
o
f
j
o
u
r
n
alis
m
an
d
AI
-
d
r
iv
e
n
c
o
n
ten
t
g
en
er
atio
n
.
T
h
e
p
r
o
p
o
s
ed
al
g
o
r
ith
m
,
with
its
m
u
ltimo
d
al
c
o
m
b
in
atio
n
o
f
m
ac
h
in
e
lear
n
i
n
g
an
d
tr
a
n
s
f
o
r
m
e
r
alg
o
r
ith
m
s
,
o
f
f
e
r
s
a
p
r
o
m
is
in
g
s
o
lu
tio
n
to
th
e
ch
allen
g
e
o
f
g
en
er
atin
g
n
ews a
r
ticles f
r
o
m
i
m
ag
es.
W
ith
f
u
r
th
er
r
ef
in
em
en
t a
n
d
im
p
r
o
v
em
en
t,
th
e
alg
o
r
ith
m
h
as th
e
p
o
ten
tial to
r
ev
o
l
u
tio
n
ize
th
e
f
ield
o
f
d
i
g
ital jo
u
r
n
alis
m
.
T
h
e
p
r
o
s
p
ec
ts
o
f
th
is
r
esear
c
h
ar
e
b
r
o
a
d
an
d
s
tim
u
latin
g
.
T
h
e
alg
o
r
ith
m
,
w
h
ile
ef
f
ec
tiv
e,
p
r
esen
ts
o
p
p
o
r
tu
n
ities
f
o
r
r
e
f
in
em
en
t.
E
n
h
an
ce
m
e
n
ts
in
co
n
tex
t
u
n
d
er
s
tan
d
in
g
an
d
jo
u
r
n
alis
tic
s
ty
le
g
en
er
atio
n
c
o
u
l
d
b
e
ac
h
iev
e
d
b
y
e
x
p
lo
r
i
n
g
a
d
v
an
ce
d
tr
a
n
s
f
o
r
m
er
m
o
d
els
f
o
r
tex
t
g
e
n
er
atio
n
an
d
m
o
r
e
c
o
m
p
lex
C
NN
m
o
d
els
f
o
r
im
ag
e
ca
p
tio
n
in
g
.
Fu
r
th
e
r
,
th
e
alg
o
r
ith
m
'
s
r
ea
l
-
wo
r
ld
p
er
f
o
r
m
a
n
ce
n
ee
d
s
ad
d
itio
n
al
ex
p
lo
r
atio
n
.
Fu
tu
r
e
wo
r
k
co
u
l
d
in
clu
d
e
test
in
g
w
ith
d
iv
er
s
e
im
ag
es
an
d
in
v
ar
ied
co
n
tex
ts
.
Mo
r
eo
v
er
,
th
e
a
lg
o
r
ith
m
'
s
p
o
ten
tial
ex
ten
d
s
b
ey
o
n
d
d
ig
ital
jo
u
r
n
a
lis
m
to
o
th
er
f
ield
s
r
eq
u
ir
in
g
au
to
m
ated
im
ag
e
-
b
ased
co
n
te
n
t
g
en
er
atio
n
,
s
u
ch
as
s
o
cial
m
ed
ia
m
an
ag
em
en
t
an
d
p
u
b
lic
r
elatio
n
s
.
I
n
co
n
cl
u
s
io
n
,
th
is
r
esear
ch
o
p
en
s
v
ast
p
o
s
s
ib
ilit
ie
s
.
W
ith
f
u
r
th
er
r
ef
in
e
m
en
t,
t
h
e
alg
o
r
it
h
m
co
u
ld
s
ig
n
if
ica
n
tly
co
n
tr
i
b
u
te
to
au
to
m
ate
d
jo
u
r
n
alis
m
,
AI
-
d
r
i
v
en
c
o
n
ten
t
g
en
er
atio
n
,
an
d
o
th
er
te
x
t a
n
d
im
ag
e
-
r
elate
d
au
t
o
m
atio
n
s
ce
n
ar
io
s
.
RE
F
E
R
E
NC
E
S
[
1
]
C
.
A
.
K
.
A
.
K
o
u
n
t
a
,
B
.
K
a
ms
u
-
F
o
g
u
e
m,
F
.
N
o
u
r
e
d
d
i
n
e
,
a
n
d
F
.
T
a
n
g
a
r
a
,
“
M
u
l
t
i
m
o
d
a
l
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
p
r
e
d
i
c
t
i
n
g
t
h
e
c
h
o
i
c
e
o
f
c
u
t
p
a
r
a
met
e
r
s
i
n
t
h
e
m
i
l
l
i
n
g
p
r
o
c
e
ss,”
I
n
t
e
l
l
i
g
e
n
t
S
y
st
e
m
s
w
i
t
h
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
1
6
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
sw
a
.
2
0
2
2
.
2
0
0
1
1
2
.
[
2
]
T.
B
a
l
t
r
u
s
a
i
t
i
s,
C
.
A
h
u
j
a
,
a
n
d
L.
-
P
.
M
o
r
e
n
c
y
,
“
M
u
l
t
i
m
o
d
a
l
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
:
a
s
u
r
v
e
y
a
n
d
t
a
x
o
n
o
m
y
,
”
I
EEE
T
ra
n
sa
c
t
i
o
n
s
o
n
Pa
t
t
e
r
n
An
a
l
y
si
s
a
n
d
Ma
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
4
1
,
n
o
.
2
,
p
p
.
4
2
3
–
4
4
3
,
F
e
b
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
1
8
.
2
7
9
8
6
0
7
.
[
3
]
T.
B
a
l
t
r
u
š
a
i
t
i
s,
C
.
A
h
u
j
a
,
a
n
d
L
.
-
P
.
M
o
r
e
n
c
y
,
“
C
h
a
l
l
e
n
g
e
s a
n
d
a
p
p
l
i
c
a
t
i
o
n
s i
n
m
u
l
t
i
mo
d
a
l
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
,
”
i
n
T
h
e
H
a
n
d
b
o
o
k
o
f
Mu
l
t
i
m
o
d
a
l
-
M
u
l
t
i
s
e
n
s
o
r
I
n
t
e
r
f
a
c
e
s:
F
o
u
n
d
a
t
i
o
n
s,
U
ser
M
o
d
e
l
i
n
g
,
a
n
d
C
o
m
m
o
n
M
o
d
a
l
i
t
y
C
o
m
b
i
n
a
t
i
o
n
s
-
Vo
l
u
m
e
2
,
A
ss
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
i
n
g
M
a
c
h
i
n
e
r
y
,
2
0
1
8
,
p
p
.
1
7
–
4
8
.
[
4
]
P
.
P
.
Li
a
n
g
,
A
.
Za
d
e
h
,
a
n
d
L
.
P
.
M
o
r
e
n
c
y
,
“
F
o
u
n
d
a
t
i
o
n
s
a
n
d
t
r
e
n
d
s
i
n
m
u
l
t
i
m
o
d
a
l
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
:
p
r
i
n
c
i
p
l
e
s,
c
h
a
l
l
e
n
g
e
s,
a
n
d
o
p
e
n
q
u
e
st
i
o
n
s,
”
a
rXi
v
p
re
p
ri
n
t
a
rXi
v
:
2
2
0
9
.
0
3
4
3
0
,
2
0
2
2
.
[
5
]
T.
B
r
o
w
n
e
t
a
l
.
,
“
L
a
n
g
u
a
g
e
m
o
d
e
l
s
a
r
e
f
e
w
-
s
h
o
t
l
e
a
r
n
e
r
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
p
ro
c
e
ss
i
n
g
syst
e
m
s
,
v
o
l
.
3
3
,
p
p
.
1
8
7
7
–
1
9
0
1
,
2
0
2
0
.
[
6
]
R
.
K
o
n
c
e
l
-
K
e
d
z
i
o
r
sk
i
,
D
.
B
e
k
a
l
,
Y
.
Lu
a
n
,
M
.
La
p
a
t
a
,
a
n
d
H
.
H
a
j
i
s
h
i
r
z
i
,
“
Te
x
t
g
e
n
e
r
a
t
i
o
n
f
r
o
m
k
n
o
w
l
e
d
g
e
g
r
a
p
h
s
w
i
t
h
g
r
a
p
h
t
r
a
n
sf
o
r
mers
,
”
a
rX
i
v
p
re
p
r
i
n
t
a
rX
i
v
:
1
9
0
4
.
0
2
3
4
2
,
2
0
1
9
.
[
7
]
M
.
Z.
A
l
o
m
e
t
a
l
.
,
“
A
s
t
a
t
e
-
of
-
t
h
e
-
a
r
t
su
r
v
e
y
o
n
d
e
e
p
l
e
a
r
n
i
n
g
t
h
e
o
r
y
a
n
d
a
r
c
h
i
t
e
c
t
u
r
e
s
,
”
El
e
c
t
ro
n
i
c
s
,
v
o
l
.
8
,
n
o
.
3
,
M
a
r
.
2
0
1
9
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s
8
0
3
0
2
9
2
.
[
8
]
K
.
A
r
u
l
k
u
m
a
r
a
n
,
M
.
P
.
D
e
i
s
e
n
r
o
t
h
,
M
.
B
r
u
n
d
a
g
e
,
a
n
d
A
.
A
.
B
h
a
r
a
t
h
,
“
D
e
e
p
r
e
i
n
f
o
r
c
e
m
e
n
t
l
e
a
r
n
i
n
g
:
a
b
r
i
e
f
su
r
v
e
y
,
”
I
E
EE
S
i
g
n
a
l
Pro
c
e
ssi
n
g
Ma
g
a
zi
n
e
,
v
o
l
.
3
4
,
n
o
.
6
,
p
p
.
2
6
–
3
8
,
N
o
v
.
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
M
S
P
.
2
0
1
7
.
2
7
4
3
2
4
0
.
[
9
]
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
,
”
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
p
r
o
c
e
ssi
n
g
s
y
st
e
m
s
,
v
o
l
.
3
0
,
2
0
1
7
.
[
1
0
]
J.
D
e
v
l
i
n
,
M
.
W
.
C
h
a
n
g
,
K
.
Le
e
,
a
n
d
K
.
T
o
u
t
a
n
o
v
a
,
“
B
e
r
t
:
p
r
e
-
t
r
a
i
n
i
n
g
o
f
d
e
e
p
b
i
d
i
r
e
c
t
i
o
n
a
l
t
r
a
n
sf
o
r
m
e
r
s
f
o
r
l
a
n
g
u
a
g
e
u
n
d
e
r
s
t
a
n
d
i
n
g
,
”
a
rX
i
v
p
r
e
p
r
i
n
t
a
rX
i
v
:
1
8
1
0
.
0
4
8
0
5
,
2
0
1
8
.
[
1
1
]
P
.
H
o
w
l
a
d
e
r
,
P
.
P
a
u
l
,
M
.
M
a
d
a
v
i
,
L.
B
e
w
o
o
r
,
a
n
d
V
.
S
.
D
e
sh
p
a
n
d
e
,
“
F
i
n
e
t
u
n
i
n
g
t
r
a
n
sf
o
r
mer
b
a
s
e
d
B
E
R
T
m
o
d
e
l
f
o
r
g
e
n
e
r
a
t
i
n
g
t
h
e
a
u
t
o
m
a
t
i
c
b
o
o
k
s
u
mm
a
r
y
,
”
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
n
Re
c
e
n
t
a
n
d
I
n
n
o
v
a
t
i
o
n
T
r
e
n
d
s
i
n
C
o
m
p
u
t
i
n
g
a
n
d
C
o
m
m
u
n
i
c
a
t
i
o
n
,
v
o
l
.
1
0
,
n
o
.
1
s
,
p
p
.
3
4
7
–
3
5
2
,
D
e
c
.
2
0
2
2
,
d
o
i
:
1
0
.
1
7
7
6
2
/
i
j
r
i
t
c
c
.
v
1
0
i
1
s
.
5
9
0
2
.
[
1
2
]
M
.
V
K
o
r
o
t
e
e
v
,
“
B
ER
T
:
a
r
e
v
i
e
w
o
f
a
p
p
l
i
c
a
t
i
o
n
s
i
n
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ssi
n
g
a
n
d
u
n
d
e
r
s
t
a
n
d
i
n
g
,
”
a
rX
i
v
p
r
e
p
r
i
n
t
a
rXi
v
:
2
1
0
3
.
1
1
9
4
3
,
2
0
2
1
.
[
1
3
]
T.
L
e
S
c
a
o
e
t
a
l
.
,
“
B
l
o
o
m:
A
1
7
6
b
-
p
a
r
a
met
e
r
o
p
e
n
-
a
c
c
e
ss m
u
l
t
i
l
i
n
g
u
a
l
l
a
n
g
u
a
g
e
mo
d
e
l
,
”
a
rXi
v
:
2
2
1
1
.
0
5
1
0
0
,
2
0
2
2
.
[
1
4
]
H
.
S
h
a
r
ma
,
M
.
A
g
r
a
h
a
r
i
,
S
.
K
.
S
i
n
g
h
,
M
.
F
i
r
o
j
,
a
n
d
R
.
K
.
M
i
s
h
r
a
,
“
I
mag
e
c
a
p
t
i
o
n
i
n
g
:
a
c
o
m
p
r
e
h
e
n
si
v
e
su
r
v
e
y
,
”
i
n
2
0
2
0
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
P
o
w
e
r
El
e
c
t
r
o
n
i
c
s
&
I
o
T
Ap
p
l
i
c
a
t
i
o
n
s
i
n
R
e
n
e
w
a
b
l
e
En
e
rg
y
a
n
d
i
t
s
C
o
n
t
ro
l
(
P
ARC
)
,
F
e
b
.
2
0
2
0
,
p
p
.
3
2
5
–
3
2
8
,
d
o
i
:
1
0
.
1
1
0
9
/
P
A
R
C
4
9
1
9
3
.
2
0
2
0
.
2
3
6
6
1
9
.
[
1
5
]
M
.
D
.
Z.
H
o
ssa
i
n
,
F
.
S
o
h
e
l
,
M
.
F
.
S
h
i
r
a
t
u
d
d
i
n
,
a
n
d
H
.
La
g
a
,
“
A
c
o
mp
r
e
h
e
n
s
i
v
e
s
u
r
v
e
y
o
f
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
i
mag
e
c
a
p
t
i
o
n
i
n
g
,
”
AC
M
C
o
m
p
u
t
i
n
g
S
u
rv
e
y
s
,
v
o
l
.
5
1
,
n
o
.
6
,
p
p
.
1
–
3
6
,
N
o
v
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
4
5
/
3
2
9
5
7
4
8
.
[
1
6
]
O
.
V
i
n
y
a
l
s,
A
.
To
s
h
e
v
,
S
.
B
e
n
g
i
o
,
a
n
d
D
.
Er
h
a
n
,
“
S
h
o
w
a
n
d
T
e
l
l
:
l
e
ss
o
n
s
l
e
a
r
n
e
d
f
r
o
m
t
h
e
2
0
1
5
M
S
C
O
C
O
i
mag
e
c
a
p
t
i
o
n
i
n
g
c
h
a
l
l
e
n
g
e
,
”
I
EEE
T
r
a
n
sa
c
t
i
o
n
s
o
n
P
a
t
t
e
r
n
A
n
a
l
y
s
i
s
a
n
d
Ma
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
9
,
n
o
.
4
,
p
p
.
6
5
2
–
6
6
3
,
A
p
r
.
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
1
6
.
2
5
8
7
6
4
0
.
[
1
7
]
C
.
S
z
e
g
e
d
y
,
S
.
I
o
f
f
e
,
V
.
V
a
n
h
o
u
c
k
e
,
a
n
d
A
.
A
l
e
m
i
,
“
I
n
c
e
p
t
i
o
n
-
v
4
,
I
n
c
e
p
t
i
o
n
-
R
e
sN
e
t
a
n
d
t
h
e
i
m
p
a
c
t
o
f
R
e
si
d
u
a
l
c
o
n
n
e
c
t
i
o
n
s
o
n
l
e
a
r
n
i
n
g
,
”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
AAAI
C
o
n
f
e
r
e
n
c
e
o
n
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
1
,
n
o
.
1
,
F
e
b
.
2
0
1
7
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
1
i
1
.
1
1
2
3
1
.
[
1
8
]
Y
.
Le
c
u
n
,
L.
B
o
t
t
o
u
,
Y
.
B
e
n
g
i
o
,
a
n
d
P
.
H
a
f
f
n
e
r
,
“
G
r
a
d
i
e
n
t
-
b
a
s
e
d
l
e
a
r
n
i
n
g
a
p
p
l
i
e
d
t
o
d
o
c
u
m
e
n
t
r
e
c
o
g
n
i
t
i
o
n
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
,
v
o
l
.
8
6
,
n
o
.
1
1
,
p
p
.
2
2
7
8
–
2
3
2
4
,
1
9
9
8
,
d
o
i
:
1
0
.
1
1
0
9
/
5
.
7
2
6
7
9
1
.
[
1
9
]
P
.
K
a
l
e
,
M
.
P
a
n
c
h
p
o
r
,
S
.
D
i
n
g
o
r
e
,
S
.
G
a
i
k
w
a
d
,
a
n
d
L
.
B
e
w
o
o
r
,
“
Tr
a
f
f
i
c
si
g
n
c
l
a
ssi
f
i
c
a
t
i
o
n
u
si
n
g
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
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
rn
a
l
o
f
S
c
i
e
n
t
i
f
i
c
R
e
se
a
rc
h
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
En
g
i
n
e
e
ri
n
g
a
n
d
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
p
p
.
1
–
1
0
,
N
o
v
.
2
0
2
1
,
d
o
i
:
1
0
.
3
2
6
2
8
/
C
S
EI
T2
1
7
5
4
5
.
[
2
0
]
T.
I
q
b
a
l
a
n
d
S
.
Q
u
r
e
s
h
i
,
“
T
h
e
s
u
r
v
e
y
:
t
e
x
t
g
e
n
e
r
a
t
i
o
n
m
o
d
e
l
s
i
n
d
e
e
p
l
e
a
r
n
i
n
g
,
”
J
o
u
r
n
a
l
o
f
K
i
n
g
S
a
u
d
U
n
i
v
e
rs
i
t
y
-
C
o
m
p
u
t
e
r
a
n
d
I
n
f
o
rm
a
t
i
o
n
S
c
i
e
n
c
e
s
,
v
o
l
.
3
4
,
n
o
.
6
,
p
p
.
2
5
1
5
–
2
5
2
8
,
J
u
n
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
k
s
u
c
i
.
2
0
2
0
.
0
4
.
0
0
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
4
,
No
.
3
,
J
u
n
e
2
0
2
4
:
3
4
3
4
-
3
4
4
2
3442
[
2
1
]
A
.
A
mi
n
-
N
e
j
a
d
,
J.
I
v
e
,
a
n
d
S
.
V
e
l
u
p
i
l
l
a
i
,
“
Ex
p
l
o
r
i
n
g
t
r
a
n
sf
o
r
m
e
r
t
e
x
t
g
e
n
e
r
a
t
i
o
n
f
o
r
me
d
i
c
a
l
d
a
t
a
s
e
t
a
u
g
me
n
t
a
t
i
o
n
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
T
w
e
l
f
t
h
L
a
n
g
u
a
g
e
Re
s
o
u
r
c
e
s
a
n
d
Ev
a
l
u
a
t
i
o
n
C
o
n
f
e
r
e
n
c
e
,
p
p
.
4
6
9
9
–
4
7
0
8
,
2
0
2
0
.
[
2
2
]
O
.
A
b
d
e
l
w
a
h
a
b
a
n
d
A
.
El
m
a
g
h
r
a
b
y
,
“
D
e
e
p
l
e
a
r
n
i
n
g
b
a
se
d
v
s
M
a
r
k
o
v
C
h
a
i
n
b
a
se
d
t
e
x
t
g
e
n
e
r
a
t
i
o
n
f
o
r
c
r
o
ss
d
o
m
a
i
n
a
d
a
p
t
a
t
i
o
n
f
o
r
sen
t
i
m
e
n
t
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
2
0
1
8
I
EE
E
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
I
n
f
o
rm
a
t
i
o
n
R
e
u
s
e
a
n
d
I
n
t
e
g
r
a
t
i
o
n
(
I
RI
)
,
J
u
l
.
2
0
1
8
,
p
p
.
2
5
2
–
2
5
5
,
d
o
i
:
1
0
.
1
1
0
9
/
I
R
I
.
2
0
1
8
.
0
0
0
4
6
.
[
2
3
]
D
.
M
a
r
c
h
e
g
g
i
a
n
i
a
n
d
L.
P
e
r
e
z
-
B
e
l
t
r
a
c
h
i
n
i
,
“
D
e
e
p
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
e
n
c
o
d
e
r
s
f
o
r
s
t
r
u
c
t
u
r
e
d
d
a
t
a
t
o
t
e
x
t
g
e
n
e
r
a
t
i
o
n
,
”
a
r
Xi
v
p
re
p
ri
n
t
a
rXi
v
:
1
8
1
0
.
0
9
9
9
5
,
2
0
1
8
.
[
2
4
]
E.
J.
H
u
e
t
a
l
.
, “
Lo
R
A
:
l
o
w
-
r
a
n
k
a
d
a
p
t
a
t
i
o
n
o
f
l
a
r
g
e
l
a
n
g
u
a
g
e
m
o
d
e
l
s
,
”
a
r
Xi
v
p
r
e
p
ri
n
t
a
r
Xi
v
:
2
1
0
6
.
0
9
6
8
5
,
2
0
2
1
.
[
2
5
]
Y
.
C
h
a
i
,
J
.
G
k
o
u
n
t
o
u
r
a
s,
G
.
G
.
K
o
,
D
.
B
r
o
o
k
s,
a
n
d
G
.
Y
.
W
e
i
,
“
I
n
t
2
.
1
:
t
o
w
a
r
d
s
f
i
n
e
-
t
u
n
a
b
l
e
q
u
a
n
t
i
z
e
d
l
a
r
g
e
l
a
n
g
u
a
g
e
mo
d
e
l
s
w
i
t
h
e
r
r
o
r
c
o
r
r
e
c
t
i
o
n
t
h
r
o
u
g
h
l
o
w
-
r
a
n
k
a
d
a
p
t
a
t
i
o
n
,
”
a
rX
i
v
p
r
e
p
r
i
n
t
a
rX
i
v
:
2
3
0
6
.
0
8
1
6
2
,
2
0
2
3
.
[
2
6
]
S
.
P
o
r
w
a
l
,
L.
B
e
w
o
o
r
,
a
n
d
V
.
D
e
s
h
p
a
n
d
e
,
“
Tr
a
n
sf
o
r
m
e
r
b
a
s
e
d
i
m
p
l
e
men
t
a
t
i
o
n
f
o
r
a
u
t
o
ma
t
i
c
b
o
o
k
su
m
mariz
a
t
i
o
n
,
”
a
rX
i
v
p
re
p
ri
n
t
a
rXi
v
:
2
3
0
1
.
0
7
0
5
7
,
2
0
2
3
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Abh
a
y
G
o
t
m
a
r
e
is
a
fin
a
l
y
e
a
r
st
u
d
e
n
t
a
t
th
e
De
p
a
rtme
n
t
o
f
In
f
o
rm
a
ti
o
n
Tec
h
n
o
l
o
g
y
,
Vis
h
wa
k
a
rm
a
In
stit
u
te
o
f
In
f
o
rm
a
ti
o
n
Tec
h
n
o
l
o
g
y
,
P
u
n
e
.
His
re
se
a
rc
h
in
tere
sts
li
e
in
th
e
field
s
o
f
Na
tu
ra
l
Lan
g
u
a
g
e
P
r
o
c
e
ss
in
g
a
n
d
Cy
b
e
r
S
e
c
u
rit
y
.
A
b
h
a
y
is
c
u
rre
n
tl
y
e
n
g
a
g
e
d
i
n
d
e
v
e
l
o
p
i
n
g
i
n
n
o
v
a
ti
v
e
so
lu
ti
o
n
s
i
n
th
e
se
a
re
a
s.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
a
b
h
a
y
.
2
2
0
1
0
4
8
1
@
v
ii
t.
a
c
.
i
n
fo
r
fu
rth
e
r
d
isc
u
ss
i
o
n
o
r
c
o
l
lab
o
ra
ti
o
n
o
n
h
is r
e
se
a
rc
h
in
tere
sts.
G
a
n
d
h
a
r
v
a
T
h
ite
is
a
n
u
n
d
e
rg
ra
d
u
a
te
st
u
d
e
n
t
a
t
Vish
wa
k
a
rm
a
In
stit
u
te
o
f
In
fo
rm
a
ti
o
n
Tec
h
n
o
lo
g
y
.
His
r
e
se
a
rc
h
a
n
d
a
c
a
d
e
m
ic
in
tere
sts
m
a
in
ly
i
n
c
lu
d
e
m
a
c
h
in
e
lea
rn
in
g
,
n
a
t
u
ra
l
la
n
g
u
a
g
e
p
r
o
c
e
ss
in
g
a
n
d
m
a
th
e
m
a
ti
c
s.
H
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
g
a
n
d
h
a
rv
a
.
2
2
0
1
0
5
7
7
@v
ii
t
.
a
c
.
in
.
La
x
m
i
Be
wo
o
r
h
o
ld
s
P
h
D
in
a
rti
ficia
l
in
telli
g
e
n
c
e
a
s
m
a
jo
r
a
n
d
o
p
t
imiz
a
ti
o
n
a
s
m
in
o
r.
C
u
rre
n
tl
y
sh
e
is
a
n
a
ss
o
c
iate
p
ro
fe
ss
o
r
a
t
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
E
n
g
in
e
e
rin
g
,
Vish
wa
k
a
rm
a
In
stit
u
te
o
f
In
fo
r
m
a
ti
o
n
Tec
h
n
o
lo
g
y
,
P
u
n
e
.
He
r
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
n
a
tu
ra
l
lan
g
u
a
g
e
p
r
o
c
e
ss
in
g
,
m
e
t
a
h
e
u
risti
c
s
tec
h
n
i
q
u
e
s,
d
e
e
p
lea
r
n
in
g
,
a
rti
ficia
l
in
tell
ig
e
n
c
e
,
a
lg
o
rit
h
m
s.
S
h
e
h
a
s
s
u
c
c
e
ss
fu
ll
y
c
o
m
p
lete
d
o
n
e
f
u
n
d
e
d
re
se
a
r
c
h
p
r
o
jec
ts.
5
I
n
tern
a
ti
o
n
a
l
P
a
ten
ts
a
re
g
ra
n
ted
in
h
e
r
n
a
m
e
.
S
h
e
h
a
s
p
u
b
li
s
h
e
d
m
o
re
th
a
n
3
5
r
e
se
a
rc
h
a
rti
c
les
in
re
p
u
tab
le
jo
u
r
n
a
ls an
d
c
o
n
fe
re
n
c
e
s
.
S
h
e
is
a
re
v
iew
e
r
o
f
lea
d
in
g
jo
u
rn
a
ls
o
f
S
p
rin
g
e
r,
El
se
v
ier an
d
IEE
E
.
S
h
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
la
x
m
i.
b
e
wo
o
r@
v
ii
t.
a
c
.
i
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.