I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
42
,
No
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
742
~
75
2
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
42
.i
3
.
pp
742
-
75
2
742
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Im
pro
v
ed int
erac
tivity
and a
utom
a
ted
res
po
nse
for v
isua
l questio
n ans
wering
Ng
uy
en
H
a
M
a
nh
K
ha
ng
,
Ng
uy
en
T
ua
n An
h
,
Ng
uy
en
M
inh
H
o
a
ng
,
B
ui T
ha
n
h H
un
g
D
a
t
a
S
c
i
e
n
c
e
La
b
o
r
a
t
o
r
y
,
F
a
c
u
l
t
y
o
f
I
n
f
o
r
ma
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
I
n
d
u
s
t
r
i
a
l
U
n
i
v
e
r
si
t
y
o
f
H
o
C
h
i
M
i
n
h
c
i
t
y
,
H
o
C
h
i
M
i
n
h
c
i
t
y
,
V
i
e
t
n
a
m
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
2
4
,
2
0
2
5
R
ev
is
ed
Ma
r
1
3
,
2
0
2
6
Acc
ep
ted
Ma
y
2
6
,
2
0
2
6
Visu
a
l
q
u
e
stio
n
a
n
sw
e
ri
n
g
(VQ
A)
sy
ste
m
s
h
a
v
e
m
a
d
e
su
b
sta
n
ti
a
l
p
r
o
g
re
ss
,
y
e
t
th
e
y
stil
l
fa
c
e
li
m
it
a
ti
o
n
s
i
n
h
a
n
d
li
n
g
c
o
m
p
lex
o
r
a
m
b
i
g
u
o
u
s
q
u
e
ries
a
n
d
su
p
p
o
rti
n
g
re
a
l
-
ti
m
e
i
n
tera
c
ti
o
n
d
u
e
t
o
re
li
a
n
c
e
o
n
lar
g
e
,
c
o
m
p
u
tatio
n
a
ll
y
e
x
p
e
n
siv
e
m
o
d
e
ls
t
h
a
t
in
c
re
a
se
late
n
c
y
a
n
d
re
strict
p
ra
c
ti
c
a
l
d
e
p
lo
y
m
e
n
t,
p
a
rti
c
u
larly
i
n
e
d
u
c
a
ti
o
n
a
l
c
o
n
tex
ts.
T
h
is
stu
d
y
a
ims
t
o
d
e
v
e
lo
p
a
n
e
fficie
n
t
a
n
d
i
n
tera
c
ti
v
e
VQ
A
sy
ste
m
t
h
a
t
e
n
h
a
n
c
e
s
a
n
sw
e
r
a
c
c
u
ra
c
y
wh
il
e
e
n
a
b
li
n
g
n
a
tu
ra
l
two
-
wa
y
c
o
m
m
u
n
ica
ti
o
n
wit
h
u
se
rs.
T
o
a
c
h
iev
e
t
h
is
g
o
a
l,
we
p
ro
p
o
se
a
li
g
h
twe
ig
h
t
m
u
lt
im
o
d
a
l
fra
m
e
wo
rk
b
a
se
d
o
n
p
re
-
train
e
d
v
isio
n
–
lan
g
u
a
g
e
m
o
d
e
ls
su
c
h
a
s
BLIP
a
n
d
fi
n
e
-
tu
n
in
g
T
5
,
c
o
m
b
in
e
d
wit
h
p
r
o
m
p
t
e
n
g
in
e
e
rin
g
to
imp
r
o
v
e
q
u
e
stio
n
u
n
d
e
rsta
n
d
i
n
g
a
n
d
a
n
sw
e
r
g
e
n
e
ra
ti
o
n
.
T
h
e
sy
ste
m
fu
rth
e
r
i
n
c
o
r
p
o
ra
tes
c
o
n
v
e
rsa
ti
o
n
a
l
c
o
n
tex
t
m
e
m
o
ry
a
n
d
a
fe
e
d
b
a
c
k
m
e
c
h
a
n
ism
th
a
t
g
e
n
e
ra
tes
c
larifica
ti
o
n
q
u
e
sti
o
n
s
wh
e
n
u
se
r
i
n
p
u
ts
a
re
a
m
b
ig
u
o
u
s,
th
e
re
b
y
stre
n
g
th
e
n
in
g
in
tera
c
ti
o
n
c
a
p
a
b
il
it
ies
.
Ex
p
e
ri
m
e
n
ts
a
re
c
o
n
d
u
c
ted
o
n
p
u
b
li
c
b
e
n
c
h
m
a
rk
d
a
tas
e
t
F
li
c
k
r8
k
,
u
sin
g
si
n
g
le
-
G
P
U
c
o
m
p
u
tati
o
n
a
l
se
tt
i
n
g
s
to
e
v
a
lu
a
te
a
c
c
u
ra
c
y
,
re
sp
o
n
se
late
n
c
y
,
a
n
d
in
tera
c
ti
o
n
e
ffe
c
ti
v
e
n
e
ss
.
T
h
e
e
x
p
e
rime
n
tal
re
su
lt
s
d
e
m
o
n
stra
te
th
a
t
th
e
p
ro
p
o
se
d
a
p
p
ro
a
c
h
a
c
h
iev
e
s
c
o
m
p
e
ti
ti
v
e
o
r
su
p
e
rio
r
a
c
c
u
ra
c
y
c
o
m
p
a
re
d
to
h
e
a
v
ier
b
a
se
li
n
e
m
o
d
e
ls,
w
h
il
e
sig
n
ifi
c
a
n
t
ly
re
d
u
c
i
n
g
i
n
fe
re
n
c
e
ti
m
e
a
n
d
e
n
a
b
li
n
g
re
a
l
-
ti
m
e
in
tera
c
ti
o
n
.
Th
e
m
a
in
c
o
n
tri
b
u
ti
o
n
s
o
f
th
is
w
o
rk
in
c
lu
d
e
a
li
g
h
twe
i
g
h
t
,
p
ro
m
p
t
-
d
ri
v
e
n
VQ
A
a
rc
h
it
e
c
tu
re
,
a
n
i
n
tera
c
ti
v
e
st
ra
teg
y
f
o
r
re
so
lv
in
g
a
m
b
i
g
u
o
u
s
q
u
e
ries
,
a
n
d
e
m
p
iri
c
a
l
e
v
id
e
n
c
e
t
h
a
t
e
fficie
n
t
m
o
d
e
ls
c
a
n
su
p
p
o
rt
a
c
c
u
ra
te
a
n
d
c
o
n
v
e
rsa
ti
o
n
a
l
VQ
A
fo
r
e
d
u
c
a
ti
o
n
a
n
d
o
th
e
r
re
a
l
-
wo
rld
a
p
p
li
c
a
ti
o
n
s.
K
ey
w
o
r
d
s
:
Au
to
m
atic
r
esp
o
n
s
e
I
n
ter
ac
tio
n
Mo
d
el
ev
alu
atio
n
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
Vis
u
al
q
u
esti
o
n
an
s
wer
in
g
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
B
u
i T
h
an
h
Hu
n
g
Data
Scien
ce
L
ab
o
r
ato
r
y
,
Fac
u
lty
o
f
I
n
f
o
r
m
atio
n
T
ec
h
n
o
l
o
g
y
I
n
d
u
s
tr
ial
Un
iv
er
s
ity
o
f
Ho
C
h
i M
in
h
city
Ho
C
h
i M
in
h
city
,
Vietn
am
E
m
ail:
b
u
ith
an
h
h
u
n
g
@
iu
h
.
ed
u
.
v
n
1.
I
NT
RO
D
UCT
I
O
N
I
n
th
e
co
n
tex
t
o
f
t
h
e
r
a
p
id
an
d
o
n
g
o
in
g
a
d
v
an
ce
m
en
t
o
f
ar
t
if
icial
in
tellig
en
ce
(
AI
)
,
v
is
u
a
l
q
u
esti
o
n
an
s
wer
in
g
(
VQA
)
h
as
em
er
g
e
d
as
o
n
e
o
f
th
e
m
o
s
t
p
r
o
m
is
in
g
an
d
h
ig
h
ly
in
ter
d
is
cip
lin
ar
y
r
esear
ch
ar
ea
s
.
As a
m
u
ltimo
d
al
p
ar
ad
i
g
m
,
VQA
co
m
b
in
es
co
m
p
u
ter
v
is
io
n
,
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
,
an
d
d
ee
p
lear
n
in
g
tech
n
iq
u
es
to
en
a
b
le
m
ac
h
in
es
to
p
er
ce
iv
e
v
is
u
al
s
ce
n
es,
r
ea
s
o
n
ab
o
u
t
th
ei
r
co
n
te
n
t,
an
d
g
en
er
ate
m
ea
n
in
g
f
u
l
an
s
wer
s
to
q
u
esti
o
n
s
ex
p
r
ess
ed
in
n
atu
r
al
lan
g
u
a
g
e.
T
h
is
ca
p
ab
ilit
y
r
ep
r
esen
ts
an
im
p
o
r
tan
t
s
tep
to
war
d
n
ar
r
o
win
g
th
e
g
a
p
b
etwe
en
h
u
m
an
co
g
n
itiv
e
u
n
d
er
s
tan
d
in
g
an
d
m
ac
h
in
e
v
is
u
al
p
er
ce
p
tio
n
.
VQA
s
y
s
tem
s
ar
e
d
esig
n
ed
n
o
t
o
n
ly
to
r
ec
o
g
n
iz
e
o
b
jects,
attr
i
b
u
tes,
an
d
r
elatio
n
s
h
ip
s
with
in
im
ag
es,
b
u
t
also
to
i
n
teg
r
ate
th
ese
v
is
u
al
cu
es
with
lin
g
u
is
tic
co
m
p
r
eh
e
n
s
io
n
in
o
r
d
er
t
o
p
r
o
d
u
ce
co
h
e
r
en
t
an
d
co
n
te
x
tu
ally
ap
p
r
o
p
r
iate
r
esp
o
n
s
es [
1
]
−
[
5
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
I
mp
r
o
ve
d
in
tera
ctivity
a
n
d
a
u
t
o
ma
ted
r
esp
o
n
s
e
fo
r
visu
a
l q
u
esti
o
n
a
n
s
w
erin
g
…
(
N
g
u
ye
n
Ha
Ma
n
h
K
h
a
n
g
)
743
T
h
e
p
r
ac
tical
s
ig
n
if
ican
ce
o
f
VQA
tech
n
o
lo
g
y
h
as
b
ee
n
in
cr
ea
s
in
g
ly
r
ec
o
g
n
ized
ac
r
o
s
s
v
ar
io
u
s
d
o
m
ain
s
,
in
clu
d
in
g
d
ig
ital
ed
u
ca
tio
n
,
in
tellig
en
t
t
u
to
r
in
g
s
y
s
tem
s
,
h
ea
lth
ca
r
e
an
aly
s
is
,
v
ir
tu
al
ass
is
tan
t
s
,
an
d
v
is
u
al
lear
n
in
g
en
v
i
r
o
n
m
e
n
ts
[
6
]
−
[
9
]
.
I
n
e
d
u
ca
tio
n
,
VQA
en
ab
les
in
ter
ac
tiv
e
lear
n
i
n
g
b
y
a
llo
win
g
s
tu
d
en
ts
to
ask
q
u
esti
o
n
s
ab
o
u
t
im
ag
es,
d
iag
r
am
s
,
o
r
i
n
f
o
g
r
ap
h
ics
an
d
r
ec
ei
v
e
im
m
e
d
iate
f
ee
d
b
ac
k
.
Similar
ly
,
in
ass
is
tiv
e
tech
n
o
lo
g
ies,
VQA
ca
n
s
u
p
p
o
r
t
v
is
u
ally
im
p
ai
r
ed
in
d
iv
id
u
als
b
y
d
escr
ib
in
g
v
is
u
al
s
ce
n
es
o
r
an
s
wer
in
g
q
u
esti
o
n
s
ab
o
u
t
t
h
eir
s
u
r
r
o
u
n
d
in
g
s
.
T
h
ese
ap
p
licatio
n
s
d
em
o
n
s
tr
ate
th
e
p
o
ten
tial
o
f
VQA
to
im
p
r
o
v
e
ac
ce
s
s
ib
i
lity
,
en
h
a
n
c
e
u
s
er
en
g
ag
e
m
en
t,
an
d
p
r
o
v
i
d
e
f
ast,
co
n
tex
t
-
awa
r
e
r
esp
o
n
s
es
th
at
ap
p
r
o
x
im
ate
h
u
m
an
-
lik
e
u
n
d
er
s
tan
d
in
g
[
1
0
]
−
[
1
2
]
.
Desp
ite
th
ese
ad
v
an
tag
es,
cu
r
r
en
t
VQA
s
y
s
tem
s
s
till
f
ac
e
s
ev
er
al
ch
allen
g
es
th
at
lim
it
th
eir
r
ea
l
-
wo
r
ld
d
ep
l
o
y
m
en
t.
T
h
eir
p
e
r
f
o
r
m
a
n
ce
o
f
ten
d
ec
lin
es wh
e
n
d
ea
lin
g
with
co
m
p
lex
o
r
am
b
i
g
u
o
u
s
q
u
esti
o
n
s
th
at
r
eq
u
ir
e
d
ee
p
er
r
ea
s
o
n
i
n
g
,
m
u
lti
-
s
tep
in
f
er
en
ce
,
o
r
ex
te
r
n
al
k
n
o
wled
g
e.
I
n
a
d
d
itio
n
,
r
ea
l
-
tim
e
in
ter
ac
tio
n
r
em
ain
s
d
if
f
icu
lt
d
u
e
to
th
e
h
ea
v
y
co
m
p
u
tatio
n
al
r
e
q
u
ir
e
m
en
ts
o
f
lar
g
e
n
eu
r
al
ar
ch
ite
ctu
r
es,
p
ar
ticu
lar
ly
tr
an
s
f
o
r
m
er
-
b
ased
v
is
io
n
–
lan
g
u
ag
e
m
o
d
els.
I
s
s
u
es
r
elate
d
to
s
ca
lab
ilit
y
an
d
r
eso
u
r
ce
co
n
s
u
m
p
tio
n
,
s
u
ch
as
h
ig
h
GPU
m
em
o
r
y
u
s
ag
e
a
n
d
lo
n
g
in
f
er
en
ce
tim
es,
f
u
r
th
er
r
estrict
th
e
u
s
e
o
f
VQ
A
in
lo
w
-
laten
cy
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
.
T
h
ese
lim
itatio
n
s
r
e
d
u
ce
s
y
s
tem
f
lex
ib
ilit
y
a
n
d
h
in
d
er
ap
p
licatio
n
s
th
at
r
eq
u
ir
e
r
o
b
u
s
t a
n
d
im
m
ed
iate
r
esp
o
n
s
es [
2
]
.
Ad
d
r
ess
in
g
th
ese
is
s
u
es
r
e
q
u
ir
es
m
o
r
e
ef
f
icien
t
m
o
d
e
l
ar
ch
itectu
r
es,
im
p
r
o
v
e
d
m
u
ltimo
d
al
r
ep
r
esen
tatio
n
lear
n
in
g
,
a
n
d
s
tr
o
n
g
e
r
r
ea
s
o
n
in
g
m
ec
h
a
n
is
m
s
th
at
b
alan
ce
ac
cu
r
ac
y
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
.
O
v
er
co
m
in
g
th
ese
ch
allen
g
es
co
u
ld
en
ab
le
f
u
tu
r
e
VQA
s
y
s
tem
s
to
f
u
n
ctio
n
as
in
tellig
en
t
ag
en
ts
ca
p
ab
le
o
f
s
u
p
p
o
r
tin
g
r
ea
l
-
tim
e,
h
u
m
a
n
-
lik
e
in
ter
ac
tio
n
ac
r
o
s
s
d
iv
er
s
e
v
is
u
al
an
d
lin
g
u
is
tic
co
n
tex
ts
[
4
]
,
[
5
]
.
E
ar
ly
b
en
ch
m
ar
k
s
,
s
u
ch
as
th
e
VQA
d
ataset
in
tr
o
d
u
ce
d
b
y
An
to
l
et
a
l
.
[
1
3
]
,
p
r
o
v
id
e
d
a
f
o
u
n
d
atio
n
f
o
r
s
y
s
tem
atic
e
v
alu
atio
n
b
u
t
r
ev
ea
led
s
tr
o
n
g
d
ataset
b
iases
th
at
en
co
u
r
ag
ed
s
h
allo
w
r
ea
s
o
n
in
g
.
L
ater
ap
p
r
o
ac
h
es
in
co
r
p
o
r
ated
atte
n
tio
n
-
b
ased
,
c
o
-
atten
tio
n
,
an
d
t
r
an
s
f
o
r
m
er
-
b
ased
m
o
d
els
to
i
m
p
r
o
v
e
m
u
ltimo
d
al
alig
n
m
en
t,
th
o
u
g
h
o
f
ten
at
th
e
co
s
t
o
f
in
cr
ea
s
ed
c
o
m
p
lex
ity
.
W
h
ile
lar
g
e
p
r
e
-
tr
ain
e
d
v
is
io
n
–
lan
g
u
ag
e
m
o
d
els
ac
h
iev
e
s
tr
o
n
g
ac
c
u
r
ac
y
,
p
r
i
o
r
s
tu
d
ies
n
o
te
lim
ited
f
o
c
u
s
o
n
r
ea
l
-
tim
e
in
te
r
ac
tio
n
,
d
ialo
g
u
e
-
b
ased
r
ea
s
o
n
i
n
g
,
an
d
clar
if
icatio
n
m
ec
h
a
n
is
m
s
.
As
a
r
esu
lt,
m
o
s
t
ex
is
tin
g
s
y
s
tem
s
s
till
o
p
er
ate
in
a
s
in
g
le
-
tu
r
n
an
s
wer
in
g
p
ar
ad
ig
m
with
o
u
t c
o
n
v
er
s
atio
n
al
f
ee
d
b
ac
k
[
1
4
]
,
[
1
5
]
.
T
o
ad
d
r
ess
th
ese
g
ap
s
,
th
is
s
tu
d
y
p
r
o
p
o
s
es
a
lig
h
tweig
h
t
,
p
r
o
m
p
t
-
d
r
iv
en
VQA
f
r
am
e
wo
r
k
th
at
p
r
io
r
itizes
ef
f
icien
c
y
a
n
d
in
t
er
ac
tiv
ity
.
T
h
e
ap
p
r
o
ac
h
le
v
er
ag
es
a
c
o
m
p
ac
t
B
L
I
P
-
b
ased
v
is
io
n
–
lan
g
u
a
g
e
m
o
d
el
c
o
m
b
in
e
d
with
p
r
o
m
p
t
en
g
in
e
er
in
g
to
im
p
r
o
v
e
q
u
es
tio
n
u
n
d
er
s
tan
d
in
g
a
n
d
an
s
wer
g
e
n
er
atio
n
wh
ile
r
ed
u
cin
g
c
o
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
I
n
ad
d
itio
n
,
th
e
s
y
s
tem
in
teg
r
ates
co
n
v
er
s
atio
n
al
co
n
t
ex
t
m
em
o
r
y
an
d
an
ac
tiv
e
clar
if
icatio
n
m
ec
h
an
is
m
,
en
ab
lin
g
it
to
ask
f
o
llo
w
-
u
p
q
u
esti
o
n
s
wh
en
u
s
er
in
p
u
ts
ar
e
am
b
ig
u
o
u
s
.
Un
lik
e
p
r
io
r
s
tatic
VQA
s
y
s
tem
s
,
th
e
p
r
o
p
o
s
ed
f
r
am
ew
o
r
k
s
u
p
p
o
r
ts
r
ea
l
-
tim
e,
two
-
w
ay
in
ter
ac
tio
n
an
d
ad
ap
tiv
e
r
ea
s
o
n
i
n
g
,
m
a
k
in
g
it
p
ar
ticu
lar
ly
s
u
itab
le
f
o
r
ed
u
ca
t
io
n
al
an
d
i
n
ter
ac
tiv
e
ap
p
licatio
n
s
.
Ou
r
ap
p
r
o
ac
h
in
cl
u
d
es:
u
s
in
g
th
e
T
5
T
P3
m
o
d
el
to
g
e
n
er
ate
q
u
esti
o
n
s
f
r
o
m
p
h
o
t
o
ca
p
tio
n
s
;
ap
p
ly
in
g
a
co
m
p
ac
t
B
L
I
P
m
o
d
el
co
m
b
in
ed
wit
h
p
r
o
m
p
t
en
g
in
ee
r
in
g
t
ec
h
n
iq
u
es
to
o
p
tim
ize
in
p
u
t
q
u
er
ies
an
d
g
en
e
r
ate
d
escr
ip
tiv
e
an
d
in
f
o
r
m
atio
n
-
r
ich
an
s
wer
s
;
b
u
ild
a
d
escr
ip
tiv
e
VQA
d
ataset
f
r
o
m
Fli
ck
r
8
k
an
d
d
esig
n
a
r
eso
u
r
ce
-
o
p
tim
ized
tr
ain
in
g
p
r
o
ce
s
s
.
Key
co
n
tr
ib
u
tio
n
s
to
th
e
s
tu
d
y
in
clu
d
e:
−
Au
to
m
at
ically
g
en
er
ate
q
u
esti
o
n
s
u
s
in
g
T
5
T
P3
f
r
o
m
im
a
g
e
ca
p
tio
n
s
,
p
r
o
d
u
cin
g
d
i
v
er
s
e
a
n
d
co
n
tex
tu
ally
ap
p
r
o
p
r
iate
q
u
esti
o
n
s
f
o
r
v
is
u
al
co
n
ten
t.
−
I
n
teg
r
ate
lig
h
tweig
h
t
B
L
I
P
wi
th
p
r
o
m
p
t
en
g
in
ee
r
in
g
to
i
m
p
r
o
v
e
th
e
ac
cu
r
ac
y
,
c
o
h
er
en
ce
,
an
d
d
escr
ip
tiv
e
q
u
ality
o
f
g
en
er
a
te
d
an
s
wer
s
.
−
B
u
ild
a
d
escr
ip
tiv
e
VQA
d
at
aset
f
r
o
m
Fli
ck
r
8
k
,
wh
er
e
an
s
wer
s
co
r
r
esp
o
n
d
to
f
u
ll
ca
p
tio
n
s
,
en
ab
lin
g
r
ich
er
r
esp
o
n
s
es th
an
tr
ad
itio
n
al
VQA
d
atasets
.
−
Use
a
r
eso
u
r
ce
-
ef
f
icien
t
tr
ain
in
g
an
d
ev
alu
atio
n
f
r
am
ew
o
r
k
,
in
clu
d
i
n
g
f
r
ee
zin
g
th
e
v
is
io
n
en
co
d
e
r
,
g
r
ad
ien
t
ac
cu
m
u
latio
n
,
f
p
1
6
o
p
tim
izatio
n
,
an
d
m
etr
ics
s
u
ch
as
m
ea
n
q
u
esti
o
n
s
im
ilar
ity
,
m
ea
n
q
u
esti
o
n
–
ca
p
tio
n
s
im
ilar
ity
,
u
n
i
q
u
e
q
u
e
s
tio
n
r
atio
,
B
L
E
U,
an
d
R
OUGE
.
I
n
ad
d
itio
n
to
th
e
in
tr
o
d
u
ctio
n
,
th
e
r
em
ain
d
er
o
f
th
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
ws.
Par
t
2
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
p
r
o
v
id
es
a
d
etailed
an
aly
s
is
o
f
its
in
d
iv
id
u
al
co
m
p
o
n
en
ts
.
Par
t
3
d
escr
ib
es
th
e
ex
p
er
im
en
tal
s
etu
p
an
d
r
ep
o
r
t
s
a
co
m
p
a
r
ativ
e
e
v
alu
atio
n
o
f
o
u
r
ap
p
r
o
ac
h
ag
ai
n
s
t
ex
is
tin
g
m
eth
o
d
s
.
Fin
ally
,
Par
t 4
s
u
m
m
ar
izes th
e
m
ain
c
o
n
clu
s
i
o
n
s
o
f
t
h
is
s
tu
d
y
an
d
d
i
s
cu
s
s
es p
o
ten
tial d
ir
ec
tio
n
s
f
o
r
f
u
tu
r
e
r
esear
ch
.
2.
M
E
T
H
O
D
2
.
1
.
T
he
pro
po
s
ed
m
et
ho
d
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
is
d
esig
n
ed
as
a
two
-
s
tag
e
f
r
am
ewo
r
k
th
at
in
teg
r
ates
b
o
th
au
to
m
ati
c
q
u
esti
o
n
g
en
er
atio
n
a
n
d
in
tellig
en
t
a
n
s
wer
p
r
ed
ictio
n
,
en
ab
lin
g
a
m
o
r
e
s
ea
m
less
an
d
co
n
tex
tu
ally
g
r
o
u
n
d
ed
in
ter
ac
tio
n
b
etwe
en
v
is
u
al
u
n
d
er
s
tan
d
in
g
an
d
lan
g
u
ag
e
r
ea
s
o
n
in
g
.
Sp
ec
if
ically
,
th
e
f
r
am
ewo
r
k
o
p
er
ates
th
r
o
u
g
h
two
p
r
im
ar
y
p
h
ases
:
i)
Au
to
m
atic
q
u
esti
o
n
g
en
er
atio
n
f
r
o
m
im
a
g
es,
wh
er
e
v
is
u
al
in
p
u
ts
ar
e
p
r
o
ce
s
s
ed
to
p
r
o
d
u
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
42
,
No
.
3
,
J
u
n
e
20
2
6
:
7
4
2
-
75
2
744
lin
g
u
is
tically
d
iv
er
s
e
an
d
s
em
an
tically
co
h
er
en
t
q
u
esti
o
n
s
th
at
r
ef
lect
th
e
k
ey
elem
en
ts
,
o
b
jects,
an
d
r
elatio
n
s
h
ip
s
with
in
th
e
im
ag
e
;
an
d
ii)
Qu
esti
o
n
an
s
wer
in
g
b
ased
o
n
im
ag
e
c
o
n
ten
t,
in
wh
i
ch
th
e
g
en
e
r
ated
o
r
u
s
er
-
p
r
o
v
id
ed
q
u
esti
o
n
is
an
aly
ze
d
an
d
an
s
wer
ed
u
s
in
g
a
f
in
e
-
tu
n
ed
B
L
I
P
m
o
d
el
[
1
6
]
i
n
co
n
ju
n
ctio
n
with
p
r
o
m
p
t e
n
g
in
ee
r
in
g
to
en
s
u
r
e
co
n
tex
tu
al
ac
cu
r
ac
y
a
n
d
h
u
m
a
n
-
lik
e
f
lu
e
n
cy
.
Fig
u
r
e
1
p
r
esen
ts
th
e
o
v
e
r
all
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
wh
ich
in
v
o
lv
e
s
s
ev
er
al
in
ter
co
n
n
e
cted
s
tep
s
th
at
b
r
i
d
g
e
v
is
u
al
p
er
ce
p
tio
n
an
d
n
at
u
r
al
lan
g
u
ag
e
r
ea
s
o
n
in
g
.
T
h
e
s
y
s
tem
f
ir
s
t
ex
tr
ac
ts
h
ig
h
-
lev
el
v
is
u
al
r
e
p
r
esen
tatio
n
s
f
r
o
m
th
e
in
p
u
t
im
a
g
e,
t
h
en
lev
er
ag
es
th
e
T
5
T
P3
[
1
7
]
b
ased
q
u
esti
o
n
g
en
er
atio
n
m
o
d
u
le
to
a
u
to
m
at
ically
f
o
r
m
u
late
r
ele
v
an
t
q
u
es
tio
n
s
.
Su
b
s
eq
u
e
n
tly
,
t
h
e
f
in
e
-
t
u
n
ed
B
L
I
P
m
o
d
el,
g
u
id
ed
b
y
ca
r
ef
u
lly
c
r
af
ted
p
r
o
m
p
ts
,
in
ter
p
r
ets
b
o
th
th
e
im
a
g
e
f
ea
tu
r
es
an
d
th
e
tex
tu
al
q
u
er
y
to
g
en
er
ate
a
n
an
s
wer
th
at
alig
n
s
with
th
e
v
is
u
al
co
n
tex
t.
T
h
is
p
ip
elin
e
n
o
t
o
n
ly
en
h
an
ce
s
th
e
d
ep
t
h
an
d
d
iv
e
r
s
i
ty
o
f
q
u
esti
o
n
–
an
s
wer
p
air
s
b
u
t
a
ls
o
im
p
r
o
v
es
s
y
s
tem
ad
a
p
ta
b
ilit
y
ac
r
o
s
s
d
i
f
f
er
en
t
d
o
m
ai
n
s
an
d
in
ter
ac
tio
n
s
ce
n
ar
io
s
.
T
h
e
d
etailed
p
r
o
ce
d
u
r
e
o
f
ea
ch
s
tag
e
is
d
escr
ib
ed
as f
o
llo
ws.
Fig
u
r
e
1
.
Ov
e
r
v
iew
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
Data
p
r
e
-
p
r
o
ce
s
s
in
g
:
I
m
ag
es
an
d
ca
p
tio
n
s
f
r
o
m
th
e
Fli
ck
r
8
k
d
ataset
ar
e
u
s
ed
as
in
p
u
t.
C
ap
tio
n
s
ar
e
s
tan
d
ar
d
ized
,
an
d
im
ag
es
a
r
e
co
n
v
er
ted
in
to
a
f
o
r
m
at
s
u
itab
le
f
o
r
th
e
v
is
u
al
m
o
d
el.
Au
to
m
atic
q
u
esti
o
n
g
en
er
atio
n
u
s
in
g
T
5
T
P3
:
T
h
e
T
5
T
P3
m
o
d
el
tak
es
im
ag
e
ca
p
tio
n
s
as
in
p
u
t
an
d
g
en
er
ates
r
elev
an
t
q
u
esti
o
n
s
,
cr
ea
tin
g
co
n
tex
tu
al
q
u
esti
o
n
–
an
s
wer
p
air
s
th
at
r
ef
lect
th
e
im
ag
e
co
n
ten
t.
Descr
ip
tiv
e
VQA
d
ataset
co
n
s
tr
u
ctio
n
:
I
m
ag
es,
g
e
n
er
at
ed
q
u
esti
o
n
s
,
a
n
d
o
r
ig
in
al
ca
p
tio
n
s
ar
e
c
o
m
b
in
e
d
to
f
o
r
m
tr
ain
in
g
s
am
p
les
(
im
ag
e,
q
u
esti
o
n
,
a
n
s
wer
)
.
An
s
wer
s
ar
e
m
ain
tain
ed
as
d
etailed
d
escr
ip
tio
n
s
r
ath
er
th
an
th
e
s
h
o
r
t
r
esp
o
n
s
es
ty
p
ical
o
f
tr
ad
itio
n
al
VQA
d
atasets
.
B
L
I
P
-
V
QA
m
o
d
el
tr
ain
in
g
:
B
L
I
P
-
VQA
is
u
s
ed
as
th
e
q
u
esti
o
n
-
an
s
wer
in
g
m
o
d
el.
T
h
e
Vis
io
n
E
n
co
d
er
is
f
r
o
ze
n
,
wh
ile
g
r
ad
i
en
t
ac
cu
m
u
latio
n
an
d
FP
1
6
ar
e
ap
p
lied
to
r
ed
u
ce
co
m
p
u
tatio
n
al
co
s
t.
Pro
m
p
t
e
n
g
in
ee
r
in
g
is
u
s
ed
d
u
r
in
g
tr
a
in
in
g
a
n
d
i
n
f
er
en
ce
to
en
co
u
r
ag
e
co
h
er
en
t
an
d
co
m
p
lete
an
s
wer
s
.
Per
f
o
r
m
an
ce
ev
alu
atio
n
:
Qu
esti
o
n
q
u
ali
ty
is
as
s
ess
ed
u
s
in
g
Me
an
Q
u
esti
o
n
Similar
ity
,
Me
an
Qu
esti
o
n
–
C
ap
tio
n
Simi
lar
ity
,
an
d
Un
i
q
u
e
Qu
esti
o
n
R
atio
.
An
s
wer
q
u
ality
is
ev
alu
a
ted
u
s
in
g
B
L
E
U
-
n
an
d
R
OUGE
-
L
,
en
ab
lin
g
a
co
m
p
r
eh
en
s
iv
e
ass
ess
m
en
t
o
f
b
o
th
th
e
d
ataset
an
d
m
o
d
el
p
er
f
o
r
m
an
ce
.
T
h
e
ab
o
v
e
p
r
o
ce
s
s
en
s
u
r
es
th
at
th
e
m
o
d
el
is
b
o
th
ca
p
ab
le
o
f
a
u
to
m
at
ically
g
en
er
atin
g
c
o
n
tex
tu
al
VQA
tr
ain
in
g
d
ata,
o
p
tim
izin
g
t
r
ain
in
g
an
d
in
f
e
r
en
ce
f
o
r
a
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
ir
o
n
m
e
n
t,
a
n
d
p
r
o
v
id
i
n
g
a
n
s
wer
s
th
at
ar
e
d
escr
ip
tiv
e
an
d
cl
o
s
e
to
n
atu
r
a
l la
n
g
u
ag
e.
W
e
will
d
escr
ib
e
ea
ch
p
ar
t i
n
d
etail
in
th
e
n
ex
t
s
ess
io
n
.
2
.
2
.
G
ener
a
t
e
qu
estio
n
-
a
ns
wer
pa
irs f
ro
m
ima
g
es
T
h
e
m
o
d
el
p
r
o
p
o
s
ed
in
th
is
s
t
u
d
y
is
d
esig
n
ed
to
en
h
an
ce
b
o
th
th
e
q
u
ality
an
d
co
n
tex
t
u
al
r
elev
an
ce
o
f
q
u
esti
o
n
s
an
d
a
n
s
wer
s
with
in
a
VQA
f
r
am
ewo
r
k
.
Un
lik
e
co
n
v
e
n
tio
n
al
s
y
s
tem
s
t
h
at
r
ely
s
o
lely
o
n
p
r
e
-
d
ef
in
ed
q
u
esti
o
n
–
an
s
wer
p
ai
r
s
,
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
i
n
teg
r
ates
th
r
ee
co
m
p
lem
en
ta
r
y
co
m
p
o
n
en
ts
th
e
T
5
T
P3
q
u
esti
o
n
g
en
er
atio
n
m
o
d
el,
a
f
in
e
-
tu
n
e
d
B
L
I
P
m
o
d
el,
an
d
p
r
o
m
p
t
en
g
in
ee
r
in
g
te
ch
n
iq
u
es
to
f
o
r
m
a
co
h
esiv
e
an
d
a
d
ap
tiv
e
a
r
ch
itec
tu
r
e.
T
h
r
o
u
g
h
th
is
in
teg
r
atio
n
,
th
e
s
y
s
tem
g
en
er
ates
s
em
an
ticall
y
r
ich
q
u
esti
o
n
s
f
r
o
m
v
is
u
al
in
p
u
ts
an
d
p
r
o
d
u
ce
s
ac
cu
r
ate,
co
n
te
x
t
-
aw
ar
e,
h
u
m
an
-
lik
e
an
s
wer
s
b
y
le
v
er
ag
in
g
m
u
ltimo
d
al
alig
n
m
e
n
t
b
etwe
en
tex
tu
al
an
d
v
is
u
al
f
ea
tu
r
es.
T
h
e
T
5
T
P
3
co
m
p
o
n
en
t
en
s
u
r
es
lin
g
u
is
tic
d
iv
er
s
ity
an
d
g
r
am
m
atica
l
f
l
u
en
cy
in
g
e
n
er
ated
q
u
esti
o
n
s
,
wh
ile
th
e
f
in
e
-
tu
n
ed
B
L
I
P
m
o
d
el
s
tr
en
g
t
h
en
s
v
is
u
al
–
tex
tu
al
r
ea
s
o
n
in
g
an
d
s
em
an
tic
co
h
er
en
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
I
mp
r
o
ve
d
in
tera
ctivity
a
n
d
a
u
t
o
ma
ted
r
esp
o
n
s
e
fo
r
visu
a
l q
u
esti
o
n
a
n
s
w
erin
g
…
(
N
g
u
ye
n
Ha
Ma
n
h
K
h
a
n
g
)
745
b
etwe
en
im
ag
e
u
n
d
er
s
tan
d
in
g
an
d
lan
g
u
ag
e
o
u
tp
u
t.
Pro
m
p
t
en
g
in
ee
r
i
n
g
f
u
r
th
er
r
e
f
in
es
m
o
d
el
b
eh
a
v
io
r
,
en
ab
lin
g
th
e
s
y
s
tem
t
o
a
d
ap
t
to
d
if
f
er
e
n
t
co
n
tex
ts
,
q
u
esti
o
n
ty
p
es,
an
d
r
esp
o
n
s
e
s
ty
le
s
with
o
u
t
ex
te
n
s
iv
e
r
etr
ain
in
g
.
T
h
e
m
ain
f
ea
tu
r
es
o
f
th
is
m
o
d
u
le
in
clu
d
e:
Au
to
m
atic
q
u
esti
o
n
g
e
n
er
atio
n
f
r
o
m
im
ag
es
:
−
T
h
e
T
5
T
P3
m
o
d
el
g
e
n
er
ates q
u
esti
o
n
s
b
as
ed
o
n
im
a
g
e
ca
p
ti
o
n
s
f
r
o
m
th
e
Fli
ck
r
8
k
d
ataset.
−
Gen
er
ated
q
u
esti
o
n
s
p
r
o
v
id
e
d
iv
er
s
e
an
d
m
ea
n
in
g
f
u
l
c
o
n
tex
ts
th
at
h
elp
th
e
m
o
d
e
l
u
tili
ze
im
ag
e
in
f
o
r
m
atio
n
ef
f
ec
tiv
ely
.
C
o
n
tex
t
-
r
ich
d
escr
ip
tiv
e
VQA
d
ataset
:
−
An
s
wer
s
ar
e
d
etailed
d
escr
ip
t
iv
e
ca
p
tio
n
s
r
ath
er
th
an
s
h
o
r
t
r
esp
o
n
s
es,
allo
win
g
th
e
s
y
s
tem
to
p
r
o
d
u
ce
m
o
r
e
in
f
o
r
m
ativ
e
o
u
tp
u
ts
.
−
Data
is
p
r
ep
r
o
ce
s
s
ed
an
d
o
r
g
a
n
ized
as a
Py
T
o
r
c
h
d
ataset
to
f
ac
ilit
ate
tr
ain
in
g
.
E
f
f
icien
t BLI
P
-
VQA
tr
ain
in
g
:
−
T
h
e
v
is
io
n
en
c
o
d
er
is
f
r
o
ze
n
t
o
r
ed
u
ce
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
−
Gr
ad
ien
t a
cc
u
m
u
latio
n
an
d
FP
1
6
ar
e
a
p
p
lied
to
o
p
tim
ize
m
e
m
o
r
y
u
s
ag
e
an
d
t
r
ain
in
g
tim
e.
−
Pro
m
p
t e
n
g
i
n
ee
r
in
g
is
in
co
r
p
o
r
ated
to
im
p
r
o
v
e
a
n
s
wer
co
h
er
en
ce
an
d
q
u
ality
.
Mu
lti
-
cr
iter
ia
ev
alu
atio
n
:
−
Qu
esti
o
n
q
u
ality
is
ass
es
s
ed
u
s
in
g
m
ea
n
q
u
esti
o
n
s
im
ilar
ity
,
m
ea
n
q
u
esti
o
n
–
ca
p
tio
n
s
im
ilar
ity
,
an
d
u
n
iq
u
e
q
u
esti
o
n
r
atio
.
−
An
s
wer
q
u
ality
is
ev
alu
ated
u
s
in
g
B
L
E
U
-
n
an
d
R
OUGE
-
L
to
m
ea
s
u
r
e
b
o
th
ac
cu
r
ac
y
an
d
co
m
p
leten
ess
.
T
h
is
ap
p
r
o
ac
h
en
a
b
les
tr
ain
in
g
an
d
d
ep
lo
y
m
en
t
in
r
eso
u
r
ce
-
lim
ited
en
v
ir
o
n
m
en
ts
wh
ile
m
ain
tain
i
n
g
th
e
ab
ilit
y
to
g
en
er
ate
d
escr
ip
t
iv
e
q
u
esti
o
n
s
an
d
an
s
wer
s
,
m
ak
in
g
it
s
u
itab
le
f
o
r
a
p
p
licatio
n
s
s
u
ch
as
lear
n
in
g
ass
is
tan
ts
an
d
v
is
u
al
ac
ce
s
s
ib
il
ity
s
y
s
tem
s
.
T
h
e
d
ata
f
lo
w
is
illu
s
tr
ated
in
Fig
u
r
e
2
an
d
in
clu
d
es th
e
f
o
llo
win
g
s
tep
s
:
−
I
m
ag
e
an
d
ca
p
tio
n
→
q
u
esti
o
n
g
en
er
atio
n
(
T
5
T
P3
)
.
−
C
o
n
s
tr
u
ctio
n
o
f
a
d
escr
ip
tiv
e
VQA
d
ataset
(
im
ag
e,
q
u
esti
o
n
,
an
s
wer
)
.
−
Fin
e
-
tu
n
in
g
B
L
I
P
-
VQA
with
p
r
o
m
p
t e
n
g
in
ee
r
in
g
.
−
I
n
p
u
t: im
ag
e
(
ViT
)
an
d
to
k
e
n
i
ze
d
q
u
esti
o
n
.
−
Ou
tp
u
t: d
escr
ip
tiv
e
an
s
wer
.
Fig
u
r
e
2
.
Data
f
l
o
w
d
escr
ip
tio
n
Pro
m
p
tin
g
tech
n
iq
u
es
:
Du
r
in
g
th
e
tr
ain
in
g
an
d
ass
ess
m
en
t
p
r
o
ce
s
s
,
we
f
o
llo
wed
p
r
o
m
p
t
m
eth
o
d
s
p
r
esen
ted
in
[
1
8
]
,
[
1
9
]
to
d
o
p
r
o
m
p
tin
g
tech
n
iq
u
es
to
im
p
r
o
v
e
th
e
q
u
ality
o
f
q
u
esti
o
n
s
an
d
an
s
wer
s
.
Qu
esti
o
n
g
en
er
atio
n
was
p
er
f
o
r
m
ed
u
s
in
g
th
e
T
5
T
P3
m
o
d
el
f
i
n
e
-
tu
n
ed
o
n
th
e
v
al
h
alla/t5
-
b
ase
-
qg
-
h
l
m
o
d
el,
with
th
e
in
p
u
t
p
r
o
m
p
t
b
ein
g
ca
p
tio
n
s
d
escr
ib
in
g
im
ag
es
f
r
o
m
th
e
Fli
ck
r
8
k
ep
is
o
d
e
.
W
e
u
s
e
p
r
o
m
p
ts
f
o
r
b
o
th
tr
ain
in
g
an
d
test
in
g
.
T
h
e
p
r
o
m
p
t
lo
o
k
s
lik
e
"Yo
u
ar
e
a
d
escr
ip
ti
v
e
VQA
ass
i
s
tan
t.
Qu
esti
o
n
:
{q
}
".
T
h
e
q
u
esti
o
n
will
b
e
r
ef
o
r
m
atted
with
th
e
p
r
o
m
p
t.
Fo
r
ex
am
p
le,
Qu
esti
o
n
:
"Wh
o
is
g
o
in
g
in
to
a
wo
o
d
en
b
u
ild
in
g
?"
will
r
esu
lt
in
"Yo
u
ar
e
a
d
escr
ip
tiv
e
VQA
ass
is
tan
t.
Qu
esti
o
n
:
W
h
o
is
g
o
i
n
g
in
to
a
wo
o
d
en
b
u
ild
in
g
?"
.
T
h
e
au
to
m
atica
lly
lab
eled
q
u
esti
o
n
s
an
d
an
s
wer
s
ar
e
th
en
u
s
ed
to
tr
ain
th
e
B
L
I
P
m
o
d
el
in
an
im
ag
e
→
p
r
o
ce
s
s
o
r
(
im
ag
e,
q
u
esti
o
n
)
→
an
s
wer
m
o
d
el.
T
h
an
k
s
to
th
e
ab
o
v
e
im
p
r
o
v
e
m
en
ts
,
th
e
s
y
s
tem
n
o
t
o
n
l
y
in
cr
ea
s
es
ac
cu
r
ac
y
b
u
t
also
im
p
r
o
v
es
t
h
e
q
u
ality
o
f
th
e
u
s
er
ex
p
er
ien
c
e
th
r
o
u
g
h
n
atu
r
al
la
n
g
u
a
g
e
i
n
ter
ac
tio
n
,
i
n
f
o
r
m
ativ
e
d
escr
i
p
tio
n
s
,
an
d
f
lex
ib
le
r
esp
o
n
s
es.
T
h
is
is
a
s
tep
a
wa
y
f
r
o
m
th
e
tr
ad
itio
n
al
VQA
m
o
d
el
to
a
m
o
r
e
d
escr
ip
tiv
e
an
d
h
u
m
a
n
e
v
is
u
al
Q&
A
s
y
s
tem
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
42
,
No
.
3
,
J
u
n
e
20
2
6
:
7
4
2
-
75
2
746
2
.
3
.
M
o
del
f
ine
-
t
un
ing
I
n
th
is
s
ec
tio
n
,
we
d
escr
ib
e
t
h
e
p
r
o
ce
s
s
o
f
a
n
s
wer
in
g
q
u
esti
o
n
s
f
r
o
m
im
a
g
e
c
o
n
ten
t
u
s
in
g
th
e
f
in
e
-
tu
n
ed
B
L
I
P
m
o
d
el
with
p
r
o
m
p
t
en
g
i
n
ee
r
in
g
.
T
h
e
s
y
s
tem
u
s
es
B
L
I
P
-
VQA
as
th
e
co
r
e
m
o
d
u
le
to
f
u
s
e
v
is
u
al
an
d
tex
tu
al
in
f
o
r
m
atio
n
.
B
L
I
P
is
ch
o
s
en
f
o
r
its
co
m
p
ac
t
y
et
ef
f
ec
tiv
e
en
d
-
to
-
e
n
d
ar
c
h
itectu
r
e,
en
ab
lin
g
im
ag
e
–
t
ex
t
alig
n
m
e
n
t,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
a
n
s
wer
g
en
er
atio
n
i
n
a
u
n
if
ied
m
o
d
el.
W
ith
p
r
o
m
p
t
en
g
in
ee
r
in
g
,
th
e
m
o
d
el
b
etter
ad
ap
ts
to
d
if
f
er
e
n
t
q
u
est
io
n
ty
p
es
an
d
p
r
o
d
u
ce
s
m
o
r
e
d
escr
ip
tiv
e
an
d
co
n
tex
tu
ally
r
ele
v
an
t a
n
s
wer
s
.
T
h
e
f
i
n
e
-
tu
n
e
d
B
L
I
P
m
o
d
el
also
in
co
r
p
o
r
ates
atten
tio
n
m
ec
h
an
is
m
s
in
s
p
ir
ed
b
y
h
ier
a
r
ch
ical
co
-
atten
tio
n
an
d
th
e
s
tack
ed
atte
n
tio
n
n
etwo
r
k
(
SAN)
to
s
tr
en
g
th
en
i
n
ter
ac
tio
n
s
b
etwe
en
v
is
u
al
r
eg
io
n
s
an
d
tex
t
r
ep
r
esen
tatio
n
s
.
T
h
ese
m
ec
h
a
n
is
m
s
iter
ativ
ely
alig
n
im
ag
e
f
ea
tu
r
es
with
lin
g
u
is
tic
to
k
en
s
,
im
p
r
o
v
in
g
th
e
m
o
d
el’
s
ab
ilit
y
to
ca
p
tu
r
e
s
p
a
tial
d
ep
en
d
en
cies,
o
b
ject
r
elat
io
n
s
h
ip
s
,
an
d
c
r
o
s
s
-
m
o
d
al
s
e
m
an
tics
,
lead
in
g
to
m
o
r
e
ac
cu
r
ate
an
d
g
r
o
u
n
d
e
d
a
n
s
wer
s
.
Ho
wev
er
,
tr
ad
itio
n
al
atten
tio
n
-
b
ased
ap
p
r
o
ac
h
es
o
f
ten
r
el
y
o
n
lo
ca
lized
v
is
u
al
f
ea
tu
r
es
an
d
s
h
o
w
lim
itatio
n
s
in
m
u
lti
-
s
tep
r
ea
s
o
n
in
g
,
l
o
n
g
-
r
an
g
e
d
ep
e
n
d
en
c
y
m
o
d
elin
g
,
an
d
d
ee
p
er
co
n
tex
tu
al
in
f
er
e
n
ce
.
T
o
ad
d
r
ess
th
is
,
th
is
s
tu
d
y
in
te
g
r
ates
th
e
B
L
I
P
f
r
am
ewo
r
k
with
p
r
o
m
p
t
en
g
in
ee
r
in
g
to
g
u
id
e
t
h
e
m
o
d
el’
s
r
ea
s
o
n
in
g
p
r
o
ce
s
s
.
B
y
d
y
n
am
ically
r
ef
in
in
g
p
r
o
m
p
ts
d
u
r
i
n
g
in
f
er
e
n
ce
,
th
e
s
y
s
tem
f
o
cu
s
es
o
n
in
f
o
r
m
ativ
e
v
is
u
al
r
eg
io
n
s
an
d
lin
g
u
is
tic
cu
es,
r
ed
u
cin
g
am
b
i
g
u
ity
an
d
i
m
p
r
o
v
i
n
g
co
m
p
o
s
itio
n
al
r
ea
s
o
n
in
g
.
T
h
is
s
tr
ateg
y
also
en
ab
les
m
o
r
e
d
etailed
an
d
co
n
tex
t
-
awa
r
e
d
escr
ip
tio
n
s
f
o
r
co
m
p
le
x
s
ce
n
es.
Ov
e
r
all,
th
e
ap
p
r
o
ac
h
b
alan
ce
s
co
m
p
u
tatio
n
al
e
f
f
ici
en
cy
,
in
ter
p
r
etab
ilit
y
,
an
d
s
e
m
an
tic
ex
p
r
ess
iv
en
ess
,
allo
win
g
th
e
VQA
s
y
s
tem
to
p
r
o
d
u
ce
ac
cu
r
ate,
co
h
er
en
t,
an
d
in
f
o
r
m
ativ
e
a
n
s
wer
s
.
I
n
th
e
f
r
ee
ze
v
is
io
n
en
co
d
er
,
we
ap
p
ly
g
r
a
d
ien
t
ac
c
u
m
u
lati
o
n
a
n
d
FP
1
6
to
r
ed
u
ce
c
o
m
p
u
tatio
n
co
s
ts
an
d
p
r
o
m
p
t
en
g
i
n
ee
r
in
g
in
tr
ain
in
g
an
d
r
ea
s
o
n
in
g
t
o
g
u
id
e
th
e
m
o
d
el
to
g
en
e
r
ate
co
m
p
lete
an
d
c
o
h
er
e
n
t
an
s
wer
s
.
Fig
u
r
e
3
d
escr
ib
es
a
d
etailed
d
escr
ip
tio
n
o
f
th
e
f
in
e
-
tu
n
in
g
p
r
o
ce
s
s
o
f
B
L
I
P.
T
h
is
p
r
o
ce
s
s
in
g
in
clu
d
es
th
e
f
o
llo
win
g
s
tep
s
:
Fig
u
r
e
3
.
Deta
iled
d
escr
ip
tio
n
o
f
th
e
f
i
n
e
-
tu
n
in
g
p
r
o
ce
s
s
o
f
B
L
I
P
Fo
r
war
d
p
ass
:
W
h
en
u
s
in
g
H
u
g
g
in
g
Face
’
s
T
r
ain
er
to
f
in
e
-
t
u
n
e
B
L
I
P
f
o
r
th
e
VQA
task
,
t
h
e
f
o
r
war
d
p
ass
p
r
o
ce
ed
s
as
f
o
llo
ws.
Data
in
p
u
t:
I
m
ag
es
ar
e
lo
ad
ed
f
r
o
m
th
e
Data
L
o
ad
er
a
n
d
co
n
v
er
ted
in
t
o
p
ix
el
_
v
alu
es
ten
s
o
r
s
ac
c
o
r
d
i
n
g
to
B
L
I
P
r
e
q
u
ir
em
e
n
ts
,
wh
ile
q
u
esti
o
n
s
ar
e
to
k
en
ized
in
to
in
p
u
t_
id
s
an
d
atten
tio
n
_
m
ask
s
u
s
in
g
th
e
B
L
I
P
to
k
en
izer
.
I
m
ag
e
f
ea
tu
r
e
ex
tr
ac
tio
n
:
T
h
e
v
is
io
n
en
co
d
e
r
(
ty
p
ically
a
Vis
io
n
T
r
an
s
f
o
r
m
e
r
–
ViT
)
p
r
o
ce
s
s
es
th
e
im
ag
e
ten
s
o
r
s
to
p
r
o
d
u
ce
em
b
ed
d
i
n
g
s
th
at
ca
p
tu
r
e
h
ig
h
-
lev
el
v
is
u
al
f
ea
tu
r
es
s
u
ch
as
o
b
jects
an
d
co
lo
r
s
.
Qu
esti
o
n
f
ea
tu
r
e
e
x
tr
ac
tio
n
:
T
h
e
T
r
an
s
f
o
r
m
er
-
b
ased
tex
t
en
co
d
er
p
r
o
ce
s
s
es
th
e
to
k
en
ized
q
u
esti
o
n
an
d
g
en
er
ates
tex
t
em
b
ed
d
in
g
s
.
Mu
ltimo
d
al
f
u
s
io
n
:
C
r
o
s
s
-
atten
tio
n
lay
er
s
e
n
ab
le
in
ter
ac
tio
n
s
b
etwe
en
tex
tu
al
an
d
v
is
u
al
r
ep
r
esen
tatio
n
s
,
p
r
o
d
u
cin
g
a
c
o
n
t
ex
tu
al
m
u
ltimo
d
al
r
ep
r
esen
tatio
n
.
An
s
wer
p
r
e
d
ic
tio
n
:
T
h
e
f
u
s
ed
r
ep
r
esen
tatio
n
is
p
ass
ed
to
a
lan
g
u
ag
e
m
o
d
elin
g
h
ea
d
(
lin
ea
r
lay
er
with
s
o
f
tm
ax
)
to
co
m
p
u
t
e
to
k
en
p
r
o
b
a
b
ilit
ies,
f
r
o
m
wh
ich
th
e
f
in
al
an
s
wer
is
g
e
n
er
at
ed
.
L
o
s
s
ca
lcu
latio
n
:
T
h
e
tr
ain
e
r
ca
lls
th
e
co
m
p
u
te
_
lo
s
s
f
u
n
ctio
n
in
B
lip
Fo
r
Qu
esti
o
n
A
n
s
wer
in
g
.
Pre
d
ictio
n
s
ar
e
co
m
p
ar
ed
wit
h
g
r
o
u
n
d
-
tr
u
th
an
s
wer
s
at
th
e
to
k
en
le
v
el
u
s
in
g
c
r
o
s
s
-
en
tr
o
p
y
l
o
s
s
.
Pad
d
in
g
to
k
en
s
ar
e
m
ask
ed
s
o
th
at
lo
s
s
is
co
m
p
u
ted
o
n
ly
o
n
v
alid
t
o
k
en
s
.
T
h
e
r
esu
ltin
g
lo
s
s
r
ef
l
ec
ts
th
e
d
if
f
er
e
n
ce
b
etwe
en
p
r
ed
icted
an
d
r
ef
er
e
n
ce
an
s
wer
s
f
o
r
th
e
c
u
r
r
en
t
b
atc
h
.
B
ac
k
p
r
o
p
ag
atio
n
:
Du
r
in
g
th
e
b
ac
k
war
d
p
ass
,
Py
T
o
r
ch
co
m
p
u
tes
g
r
ad
ien
ts
o
f
th
e
lo
s
s
f
o
r
p
ar
am
eter
s
in
th
e
v
is
io
n
en
co
d
er
,
te
x
t
en
c
o
d
er
,
a
n
d
c
r
o
s
s
-
atten
tio
n
lay
e
r
s
.
T
h
e
Ad
am
W
o
p
tim
izer
u
p
d
ates
th
e
weig
h
ts
to
m
in
im
ize
th
e
lo
s
s
.
Key
h
y
p
e
r
p
ar
am
eter
s
in
cl
u
d
e
th
e
lear
n
in
g
r
ate,
wh
ich
co
n
tr
o
ls
u
p
d
ate
m
ag
n
itu
d
e,
a
n
d
weig
h
t
d
ec
ay
,
wh
ic
h
h
elp
s
p
r
e
v
en
t
o
v
er
f
itti
n
g
.
A
lear
n
in
g
-
r
a
te
s
ch
ed
u
ler
s
u
ch
as
lin
ea
r
W
ar
m
u
p
an
d
Dec
ay
is
ty
p
ically
ap
p
lied
to
g
r
ad
u
ally
in
cr
ea
s
e
th
e
lear
n
in
g
r
ate
at
th
e
s
tar
t
o
f
tr
ain
in
g
a
n
d
d
ec
r
ea
s
e
it
later
to
im
p
r
o
v
e
co
n
v
er
g
en
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
I
mp
r
o
ve
d
in
tera
ctivity
a
n
d
a
u
t
o
ma
ted
r
esp
o
n
s
e
fo
r
visu
a
l q
u
esti
o
n
a
n
s
w
erin
g
…
(
N
g
u
ye
n
Ha
Ma
n
h
K
h
a
n
g
)
747
2
.
4
.
E
v
al
ua
t
i
on
W
e
u
s
e
ev
alu
atio
n
in
d
icato
r
s
th
at
ar
e
s
u
itab
le
f
o
r
th
e
p
r
o
b
l
em
o
f
g
en
er
atin
g
au
to
m
atic
an
s
wer
s
f
o
r
VQA:
B
L
E
U
[
2
0
]
an
d
R
OUGE
-
L
[
2
1
]
.
B
L
E
U
-
1
,
B
L
E
U
-
2
,
B
L
E
U
-
3
,
B
L
E
U
-
4
:
N
-
g
r
am
-
b
ased
m
ea
s
u
r
em
en
ts
to
ass
ess
th
e
s
im
ilar
ity
b
etwe
en
th
e
r
esu
ltin
g
an
s
wer
an
d
th
e
ac
tu
al
an
s
wer
.
B
L
E
U
-
1
:
E
v
alu
atio
n
o
f
u
n
ig
r
am
d
u
p
licatio
n
,
B
L
E
U
-
2
:
Du
p
lic
ate
ass
ess
m
en
t
o
f
2
c
o
n
s
ec
u
tiv
e
wo
r
d
s
(
b
ig
r
am
)
,
B
L
E
U
-
3
,
B
L
E
U
-
4
:
E
x
p
a
n
d
s
with
3
-
an
d
4
-
wo
r
d
p
h
r
ases
(
tr
ig
r
am
,
4
-
g
r
a
m
)
th
at
h
elp
r
ef
l
ec
t
th
e
co
h
er
e
n
ce
o
f
th
e
r
esu
lt
in
g
an
s
wer
.
B
L
E
U
m
etr
ic
is
ca
lcu
lated
in
(
1
)
:
=
×
(
∑
×
l
og
=
1
)
(
1
)
wh
er
e:
N
-
g
r
am
m
atch
in
g
:
C
o
u
n
ts
th
e
n
u
m
b
er
o
f
n
-
g
r
am
s
th
at
co
in
cid
e
with
th
e
r
e
f
er
en
ce
s
en
ten
ce
;
Pre
cisi
o
n
with
ad
ju
s
tm
en
t
:
C
alcu
lates
t
h
e
n
-
g
r
am
m
atch
r
atio
an
d
ap
p
lies
clip
p
in
g
to
a
v
o
id
r
e
p
ea
tin
g
th
e
wo
r
d
f
r
au
d
;
an
d
b
r
ev
ity
p
en
alty
(
B
P):
Pen
alty
wh
en
th
e
b
ir
th
s
en
ten
ce
i
s
to
o
s
h
o
r
t
f
o
r
th
e
r
ef
e
r
en
ce
s
en
ten
ce
.
R
OUGE
-
L
m
ea
s
u
r
es
th
e
s
im
ilar
ity
b
etwe
en
th
e
g
en
e
r
ated
a
n
d
r
ef
er
e
n
c
e
tex
ts
,
u
s
in
g
th
e
lo
n
g
est
c
o
m
m
o
n
s
u
b
s
eq
u
en
ce
(
L
C
S)
to
ass
e
s
s
co
n
ten
t
m
atch
in
g
with
o
u
t
co
n
tig
u
ity
.
T
h
is
m
etr
ic
is
b
ased
o
n
th
e
r
ec
all
an
d
p
r
ec
is
io
n
o
f
L
C
S
to
m
ea
s
u
r
e
th
e
m
atch
b
etwe
en
th
e
b
ir
th
s
en
ten
ce
a
n
d
th
e
r
ef
er
en
ce
.
I
n
ad
d
itio
n
,
we
u
s
e
th
e
f
o
llo
win
g
m
ea
s
u
r
em
en
ts
to
ev
alu
ate
t
h
e
q
u
esti
o
n
g
en
e
r
ate
d
:
Me
an
q
u
esti
o
n
s
im
ilar
ity
(
M
QS)
:
Av
er
ag
e
s
im
ilar
ity
(
e.
g
.
,
co
s
in
e
s
im
ilar
ity
)
am
o
n
g
g
en
er
ated
q
u
esti
o
n
s
in
th
e
d
ataset,
m
e
asu
r
in
g
q
u
esti
o
n
d
i
v
er
s
ity
a
n
d
r
ele
v
an
ce
.
Me
an
q
u
esti
o
n
–
ca
p
tio
n
s
im
ilar
ity
(
MQ
C
S):
Av
er
ag
e
s
im
ilar
ity
b
etwe
en
ea
ch
g
en
er
ated
q
u
esti
o
n
an
d
its
o
r
ig
i
n
al
ca
p
tio
n
,
e
v
alu
atin
g
h
o
w
well
q
u
esti
o
n
s
r
ef
lect
ca
p
tio
n
co
n
ten
t.
Un
iq
u
e
q
u
esti
o
n
r
atio
(
UQR):
Per
ce
n
tag
e
o
f
u
n
i
q
u
e
(
n
o
n
-
r
ep
ea
ted
)
q
u
esti
o
n
s
am
o
n
g
all
g
en
er
ate
d
q
u
esti
o
n
s
,
in
d
icatin
g
th
e
d
iv
e
r
s
ity
o
f
th
e
q
u
esti
o
n
g
e
n
er
atio
n
s
y
s
tem
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Da
t
a
s
et
T
h
e
d
ataset
u
s
ed
in
th
e
ex
p
er
im
en
t
was
Fli
ck
r
8
k
[
2
2
]
,
wh
i
ch
co
n
s
is
ted
o
f
8
0
0
0
im
ag
es
d
ep
ictin
g
s
itu
atio
n
s
in
ev
er
y
d
ay
life
.
E
ac
h
p
h
o
to
h
as
s
ev
er
al
s
h
o
r
t
ca
p
tio
n
s
in
E
n
g
lis
h
.
T
ab
le
1
s
h
o
ws
d
ataset
s
tat
is
tic
s
.
T
o
b
u
ild
a
VQA
d
ataset
f
r
o
m
an
im
a
g
e,
we
p
r
o
ce
e
d
wi
th
th
e
f
o
llo
win
g
s
tep
s
:
C
ap
t
io
n
p
r
e
p
r
o
ce
s
s
in
g
:
d
u
p
licate
r
e
m
o
v
al,
s
tan
d
ar
d
ize
tex
t;
C
r
ea
te
a
q
u
esti
o
n
f
r
o
m
th
e
ca
p
tio
n
u
s
in
g
th
e
T
5
T
P3
q
u
esti
o
n
g
en
er
atio
n
m
o
d
el;
Ass
ig
n
th
e
an
s
wer
to
th
e
co
r
r
esp
o
n
d
in
g
ca
p
tio
n
its
elf
(
eq
u
iv
alen
t
to
th
e
d
escr
ip
tiv
e
VQA
m
o
d
el)
an
d
ea
ch
tem
p
late
in
clu
d
es:
im
ag
e,
q
u
esti
o
n
,
an
s
wer
,
an
d
is
s
av
ed
to
a
f
ile
th
at
m
ak
es
u
p
th
e
f
in
e
-
tu
n
e
d
Fli
ck
r
8
k
d
ataset.
T
ab
le
1
.
Data
s
et
s
tatis
tic
s
D
a
t
a
s
e
t
N
u
mb
e
r
Tr
a
i
n
se
t
5
6
6
3
Te
st
set
2
4
2
8
3
.
2
.
Resul
t
W
e
u
s
ed
Kag
g
le
No
teb
o
o
k
s
,
a
clo
u
d
-
b
ased
p
latf
o
r
m
p
r
o
v
id
e
d
b
y
Kag
g
le.
T
h
e
co
m
p
u
tin
g
en
v
ir
o
n
m
en
t
is
as
f
o
llo
ws:
C
PU:
I
n
tel(
R
)
Xeo
n
(
R
)
C
PU
@
2
.
0
0
GHz
;
GPU:
NVI
DI
A
T
esla
P1
0
0
-
PC
I
E
-
1
6
GB
(
VR
AM
:
1
6
,
3
8
4
MiB
≈
1
6
GB
)
;
Op
er
atin
g
Sy
s
tem
:
Ub
u
n
tu
2
2
.
0
4
.
4
L
T
S;
Py
t
h
o
n
Ver
s
io
n
:
Py
th
o
n
3
.
1
1
.
1
3
;
Key
L
ib
r
a
r
ies:
B
L
I
P,
T
5
,
PIL
,
s
p
ac
y
,
p
an
d
as,
s
ci
k
it
-
lear
n
,
t
q
d
m
,
an
d
g
c
ar
e
p
r
o
v
id
ed
with
in
t
h
e
n
o
teb
o
o
k
.
W
e
ev
al
u
ated
q
u
esti
o
n
g
en
e
r
atio
n
r
esu
lts
u
s
in
g
MQ
C
S,
MQ
S
,
an
d
UQR
m
etr
ics
b
y
co
m
p
ar
in
g
with
B
AR
T
-
B
ASE
an
d
FLAN
-
T
5
,
th
e
r
e
s
u
lts
ar
e
©
s
h
o
wn
in
T
ab
le
2
.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
q
u
esti
o
n
g
en
er
atio
n
r
esu
lts
M
e
t
r
i
c
B
A
R
T
-
B
A
S
E
T5
TP
3
F
LA
N
-
T5
M
Q
C
S
0
.
9
5
2
7
0
.
6
8
0
2
0
.
4
2
1
2
M
Q
S
(
W
i
t
h
1
0
0
0
S
a
m
p
l
e
)
0
.
1
5
9
9
0
.
2
5
7
2
0
.
3
5
2
2
UQR
0
.
9
9
3
7
0
.
9
3
2
7
0
.
3
6
5
7
T
h
e
ev
al
u
atio
n
o
f
T
5
T
P3
ac
r
o
s
s
th
r
ee
m
etr
ics
d
em
o
n
s
tr
ates
its
o
v
er
all
e
f
f
ec
tiv
en
ess
f
o
r
ca
p
tio
n
-
to
-
q
u
esti
o
n
g
e
n
er
atio
n
.
Sp
ec
if
ic
ally
,
T
5
T
P3
ac
h
iev
es
a
n
M
QC
S
s
co
r
e
o
f
0
.
6
8
0
2
,
wh
ich
ca
n
b
e
c
o
n
s
id
er
ed
m
o
d
er
ate.
T
h
is
r
esu
lt
in
d
icate
s
th
at
th
e
g
en
er
ated
q
u
esti
o
n
s
d
o
n
o
t
s
tr
ictly
f
o
llo
w
th
e
ca
p
tio
n
wo
r
d
i
n
g
b
u
t
in
s
tead
ten
d
to
ex
p
an
d
u
p
o
n
th
e
o
r
ig
in
al
co
n
ten
t
b
y
in
tr
o
d
u
cin
g
ad
d
itio
n
al
s
em
an
tic
elem
en
ts
.
T
h
e
MQ
S
s
co
r
e
o
f
0
.
2
5
7
2
r
ef
lects
a
m
o
d
er
ate
lev
el
o
f
r
ep
etitio
n
in
q
u
esti
o
n
ty
p
es,
wh
ich
ca
n
b
e
attr
ib
u
ted
to
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
42
,
No
.
3
,
J
u
n
e
20
2
6
:
7
4
2
-
75
2
748
m
o
d
el’
s
ten
d
en
c
y
to
lear
n
co
m
m
o
n
in
ter
r
o
g
ativ
e
p
atter
n
s
(
e.
g
.
,
wh
at,
wh
y
,
h
o
w)
.
Me
an
wh
ile,
T
5
T
P3
attain
s
a
h
ig
h
UQR
s
co
r
e
o
f
0
.
9
3
2
7
,
in
d
icatin
g
s
tr
o
n
g
s
em
a
n
tic
d
iv
er
s
ity
,
as
th
e
m
o
d
el
is
ca
p
ab
le
o
f
g
e
n
er
atin
g
q
u
esti
o
n
s
f
r
o
m
m
u
ltip
le
p
er
s
p
ec
tiv
es,
s
u
ch
as
s
p
atial
r
ela
tio
n
s
h
ip
s
,
p
u
r
p
o
s
e,
an
d
co
n
te
x
tu
al
in
f
o
r
m
atio
n
.
Ov
er
all,
th
ese
ev
alu
atio
n
r
esu
lts
s
u
g
g
est
th
at
T
5
T
P3
is
th
e
m
o
s
t
s
u
itab
le
m
o
d
el
f
o
r
th
e
task
o
f
g
e
n
er
atin
g
q
u
esti
o
n
s
f
r
o
m
ca
p
tio
n
s
.
C
o
m
p
ar
ed
t
o
B
AR
T
-
B
ASE,
alth
o
u
g
h
T
5
T
P3
y
ield
s
a
s
lig
h
tly
l
o
wer
M
QC
S
s
co
r
e,
th
is
ch
ar
ac
ter
is
tic
r
ef
lects
its
s
tr
en
g
th
in
n
o
t
o
n
l
y
r
e
f
o
r
m
u
l
atin
g
ca
p
tio
n
s
b
u
t
also
ex
p
an
d
in
g
t
h
eir
s
em
an
tic
co
n
ten
t
to
p
r
o
d
u
ce
m
o
r
e
in
f
o
r
m
ativ
e
an
d
ex
p
lo
r
at
o
r
y
q
u
esti
o
n
s
.
At
th
e
s
am
e
tim
e,
th
e
co
n
s
is
ten
tly
h
ig
h
UQR
s
co
r
e
d
em
o
n
s
tr
ates
th
at
th
e
m
o
d
el
m
ain
tain
s
q
u
esti
o
n
d
iv
er
s
ity
with
o
u
t
s
ac
r
if
icin
g
r
elev
an
ce
to
th
e
ca
p
tio
n
.
I
n
c
o
n
tr
ast,
B
AR
T
-
B
ASE
p
r
im
ar
ily
em
p
h
asizes
s
u
r
f
ac
e
-
l
ev
el
f
o
r
m
v
ar
iatio
n
,
wh
ile
F
L
AN
-
T
5
ten
d
s
to
g
en
er
ate
m
o
r
e
g
en
er
alize
d
a
n
d
r
e
p
etitiv
e
q
u
esti
o
n
s
.
T
h
er
ef
o
r
e,
c
o
n
s
id
er
in
g
th
e
b
alan
c
e
am
o
n
g
r
elev
a
n
ce
(
MQ
C
S),
s
em
an
tic
d
iv
er
s
ity
,
an
d
q
u
esti
o
n
u
n
iq
u
en
ess
(
UQ
R
)
,
T
5
T
P3
was
s
elec
ted
as
th
e
p
r
im
ar
y
m
o
d
el
f
o
r
th
is
s
tu
d
y
.
I
n
th
is
s
tu
d
y
,
we
f
i
n
e
-
tu
n
e
th
e
B
L
I
P
-
VQA
(
Salesfo
r
ce
/b
lip
-
vqa
-
b
ase)
m
o
d
el
o
n
th
e
Fli
ck
r
8
k
d
ataset
r
ep
r
o
ce
s
s
ed
as
VQ
A
.
W
e
u
s
e
th
e
T
5
T
P3
au
to
-
q
u
esti
o
n
g
en
er
atio
n
m
o
d
el
t
o
g
e
n
er
ate
q
u
esti
o
n
s
f
r
o
m
th
e
o
r
ig
in
al
ca
p
tio
n
,
a
n
d
th
e
n
f
i
n
e
-
tu
n
e
t
h
e
B
L
I
P
m
o
d
el
t
o
g
en
er
ate
d
escr
ip
tiv
e
an
s
wer
s
.
T
h
e
r
esu
lts
wer
e
ev
alu
ated
u
s
in
g
B
L
E
U
an
d
R
OUGE
,
s
h
o
win
g
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
s
u
p
er
io
r
to
p
r
e
v
io
u
s
m
eth
o
d
s
s
u
ch
as
B
L
I
P
-
2
[
2
3
]
,
I
n
s
tr
u
ctB
L
I
P
[
2
4
]
,
DE
iT
+
B
er
t
[
1
]
,
B
E
iT
+
GPT2
[
1
]
.
T
ab
le
3
s
h
o
ws
co
m
p
ar
is
o
n
r
esu
lts
b
etwe
en
m
eth
o
d
s
.
Af
ter
f
in
e
-
tu
n
in
g
th
e
B
L
I
P
-
T
5
T
P3
m
o
d
el
with
t
h
e
au
to
g
e
n
er
ated
d
ata
f
r
o
m
t
h
e
Fli
ck
r
8
k
s
et,
we
ev
alu
ated
t
h
e
m
o
d
el
o
n
th
e
te
s
t
s
et
an
d
o
b
tain
e
d
th
e
f
o
llo
win
g
r
esu
lts
:
B
L
E
U
-
1
s
co
r
ed
0
.
3
2
,
B
L
E
U
-
2
s
co
r
e
d
0
.
2
7
,
B
L
E
U
-
3
s
co
r
e
d
0
.
2
3
,
B
L
E
U
-
4
s
co
r
ed
0
.
1
9
,
an
d
R
OUGE
-
L
s
co
r
ed
0
.
5
2
.
T
h
ese
r
esu
lts
ar
e
r
elativ
ely
h
ig
h
er
th
a
n
th
e
p
r
e
v
io
u
s
two
m
o
d
els,
as
th
e
h
ig
h
R
OUGE
r
esu
lts
p
r
o
v
e
th
at
th
e
m
o
d
el's
an
s
wer
s
r
etain
m
ea
n
in
g
an
d
co
v
er
t
h
e
c
o
n
te
n
t
well.
T
h
e
u
n
if
o
r
m
in
c
r
ea
s
e
in
B
L
E
U
at
n
-
g
r
am
s
m
ea
n
s
th
at
th
e
m
o
d
el
n
o
t
o
n
ly
m
atch
e
s
th
e
id
ea
b
u
t
also
d
o
es
s
o
m
o
r
e
ac
cu
r
ately
in
s
tr
u
ctu
r
e
an
d
v
o
ca
b
u
lar
y
.
T
h
i
s
s
u
g
g
ests
th
at
th
e
f
in
e
-
tu
n
e
d
B
L
I
P m
o
d
el
is
m
o
r
e
lik
ely
to
p
r
o
d
u
ce
m
o
r
e
s
em
a
n
tic,
d
escr
ip
tiv
e
an
s
wer
s
th
an
th
e
o
r
ig
in
al
m
o
d
els
th
at
h
av
e
n
o
t b
ee
n
f
in
e
-
tu
n
ed
.
Fig
u
r
e
4
illu
s
tr
ates
th
e
d
is
tr
ib
u
tio
n
o
f
B
L
E
U
-
n
s
co
r
es
(
n
=
1
to
4
)
o
b
tain
e
d
b
y
th
e
B
L
I
P
m
o
d
el
a
f
ter
f
in
e
-
tu
n
in
g
o
n
th
e
c
o
n
s
tr
u
cted
VQA
d
ataset.
As
s
h
o
w
n
in
t
h
e
f
ig
u
r
e,
th
e
B
L
E
U
-
1
s
co
r
es
ex
h
ib
it
th
e
h
ig
h
est
v
alu
es
an
d
ar
e
p
r
e
d
o
m
in
a
n
tly
co
n
ce
n
tr
ate
d
in
th
e
r
an
g
e
o
f
0
.
5
–
0
.
7
,
in
d
icatin
g
th
at
th
e
m
o
d
el
is
h
ig
h
l
y
ef
f
ec
tiv
e
at
g
e
n
er
atin
g
ac
c
u
r
ate
s
in
g
le
-
wo
r
d
o
r
u
n
ig
r
am
-
lev
el
an
s
wer
s
th
at
clo
s
ely
m
atch
th
e
r
ef
e
r
en
ce
r
esp
o
n
s
es.
T
h
is
s
u
g
g
ests
s
tr
o
n
g
p
er
f
o
r
m
a
n
ce
in
r
ec
o
g
n
i
zin
g
k
e
y
v
is
u
al
en
titi
es,
attr
ib
u
tes,
an
d
s
im
p
le
co
n
ce
p
ts
with
in
i
m
a
g
es.
I
n
c
o
n
tr
ast,
B
L
E
U
-
2
an
d
B
L
E
U
-
3
s
co
r
es
s
h
o
w
a
n
o
ticea
b
le
d
ec
lin
e,
r
ef
lectin
g
a
m
o
d
er
ate
a
b
ilit
y
to
ca
p
tu
r
e
s
h
o
r
t
-
r
an
g
e
co
n
tex
tu
al
r
elatio
n
s
h
ip
s
,
s
u
ch
as
b
ig
r
am
s
a
n
d
t
r
ig
r
am
s
,
wh
er
e
lim
ited
co
m
p
o
s
itio
n
al
s
tr
u
ctu
r
e
is
r
eq
u
ir
ed
.
T
h
is
d
ec
lin
e
im
p
lies
th
a
t
wh
ile
th
e
m
o
d
el
ca
n
p
r
o
d
u
ce
lo
ca
lly
co
h
er
e
n
t
p
h
r
ases
,
its
co
n
s
is
ten
cy
d
ec
r
ea
s
es a
s
co
n
tex
tu
al
d
ep
en
d
e
n
cie
s
in
cr
ea
s
e.
No
tab
ly
,
th
e
B
L
E
U
-
4
s
co
r
es
ar
e
s
tr
o
n
g
ly
s
k
ewe
d
to
war
d
l
o
wer
v
alu
es,
with
m
o
s
t
s
am
p
les
in
th
e
0
.
1
–
0
.
3
r
an
g
e.
T
h
is
d
is
tr
ib
u
tio
n
in
d
icate
s
t
h
at
th
e
m
o
d
el
s
till
s
tr
u
g
g
les
to
g
en
er
ate
lo
n
g
er
s
eq
u
en
ce
s
th
at
clo
s
ely
m
atch
g
r
o
u
n
d
-
tr
u
th
a
n
s
wer
s
at
th
e
f
o
u
r
-
g
r
am
lev
el
.
I
n
m
an
y
ca
s
es,
g
en
er
ated
r
e
s
p
o
n
s
es
alig
n
o
n
ly
with
p
ar
tial
s
eg
m
en
ts
o
f
th
e
r
ef
er
en
ce
s
en
ten
ce
,
r
ed
u
cin
g
p
r
ec
is
io
n
f
o
r
co
m
p
le
x
g
r
am
m
a
t
ical
s
tr
u
ctu
r
es
an
d
s
em
an
tically
r
ich
d
escr
ip
tio
n
s
.
T
h
is
b
eh
av
io
r
s
u
g
g
ests
a
ten
d
en
cy
to
wa
r
d
s
im
p
lific
atio
n
wh
en
h
an
d
lin
g
lo
n
g
er
o
r
m
o
r
e
co
m
p
lex
a
n
s
wer
s
.
Ho
wev
er
,
th
is
p
atter
n
is
co
m
m
o
n
in
au
to
m
ated
tex
t
g
e
n
er
atio
n
an
d
VQA
s
y
s
tem
s
,
wh
er
e
m
o
d
els
o
f
ten
p
r
io
r
itize
s
em
an
tic
co
r
r
ec
tn
ess
an
d
v
is
u
al
r
elev
an
ce
o
v
er
ex
ac
t
lex
ical
m
atch
in
g
.
T
h
e
r
ef
o
r
e
,
lo
wer
B
L
E
U
-
4
s
co
r
es
d
o
n
o
t
n
ec
ess
ar
ily
in
d
icate
p
o
o
r
an
s
wer
q
u
ali
ty
b
u
t
r
ath
er
r
ef
lect
th
e
tr
ad
e
-
o
f
f
b
etwe
en
f
lu
en
c
y
,
s
em
an
tic
ad
eq
u
ac
y
,
an
d
s
tr
ict
n
-
g
r
am
o
v
er
la
p
in
g
e
n
er
ativ
e
e
v
alu
atio
n
m
etr
ics.
Fig
u
r
e
5
s
h
o
ws
th
e
d
is
tr
ib
u
tio
n
o
f
R
OUGE
-
L
s
co
r
es
o
n
t
h
e
test
s
et,
wh
ich
f
o
llo
ws
a
n
a
p
p
r
o
x
im
ately
n
o
r
m
al
p
atter
n
with
a
s
lig
h
t
s
k
ew
to
war
d
lo
wer
v
alu
es.
Mo
s
t
s
am
p
les
f
all
b
etwe
en
0
.
4
a
n
d
0
.
6
5
,
i
n
d
icatin
g
th
at
th
e
m
o
d
el
o
f
ten
g
e
n
er
ate
s
an
s
wer
s
w
ith
co
n
s
id
er
ab
le
o
v
er
lap
an
d
s
tr
u
ctu
r
al
s
im
ilar
ity
to
th
e
r
ef
er
en
ce
r
esp
o
n
s
es.
T
h
e
av
er
a
g
e
s
co
r
e
o
f
ab
o
u
t
0
.
5
2
(
m
a
r
k
ed
b
y
th
e
r
ed
lin
e)
s
u
g
g
ests
a
r
ea
s
o
n
ab
ly
ac
ce
p
tab
le
p
er
f
o
r
m
an
ce
lev
el
f
o
r
g
en
e
r
ativ
e
VQA
task
s
an
d
r
ef
lect
s
th
e
m
o
d
el’
s
ab
ilit
y
to
p
r
e
s
er
v
e
k
ey
s
em
an
tic
in
f
o
r
m
atio
n
an
d
s
en
ten
ce
s
tr
u
ctu
r
e.
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
r
esu
lts
b
etwe
en
m
eth
o
d
s
M
e
t
h
o
d
B
LEU
@
1
B
LEU
@
2
B
LEU
@
3
B
LEU
@
4
R
O
U
G
E
B
LI
P
-
2
0
.
2
2
0
.
1
9
0
.
1
6
0
.
1
4
0
.
4
2
I
n
st
r
u
c
t
B
LI
P
0
.
0
8
0
.
0
6
0
.
0
5
0
.
0
4
0
.
1
5
B
LI
P
-
T5
TP3
(
o
u
r
s)
0
.
3
2
0
.
2
7
0
.
2
3
0
.
1
9
0
.
5
2
D
Ei
T
+
B
e
r
t
0
.
1
7
0
.
1
0
0
.
0
7
0
.
0
5
0
.
2
6
B
Ei
T
+
G
P
T2
0
.
2
4
0
.
1
3
0
.
0
8
0
.
0
5
0
.
2
5
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
I
mp
r
o
ve
d
in
tera
ctivity
a
n
d
a
u
t
o
ma
ted
r
esp
o
n
s
e
fo
r
visu
a
l q
u
esti
o
n
a
n
s
w
erin
g
…
(
N
g
u
ye
n
Ha
Ma
n
h
K
h
a
n
g
)
749
Fig
u
r
e
4
.
B
L
E
U
-
1
t
o
B
L
E
U
-
4
s
co
r
e
d
is
tr
ib
u
tio
n
c
h
ar
t o
n
th
e
test
d
ataset
Fig
u
r
e
5
.
R
OUGE
-
L
d
is
tr
ib
u
ti
o
n
ch
ar
t
o
n
test
s
et
Ho
wev
er
,
th
e
d
is
tr
ib
u
tio
n
r
e
m
ain
s
r
elativ
ely
wid
e,
with
s
o
m
e
s
am
p
les
s
co
r
in
g
b
elo
w
0
.
2
a
n
d
o
t
h
er
s
ap
p
r
o
ac
h
in
g
0
.
8
.
T
h
is
v
ar
ia
b
ilit
y
in
d
icate
s
in
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
ac
r
o
s
s
q
u
esti
o
n
ty
p
es
a
n
d
v
is
u
al
co
n
tex
ts
.
T
h
e
m
o
d
el
p
er
f
o
r
m
s
well
o
n
s
tr
aig
h
tf
o
r
war
d
o
r
v
is
u
ally
g
r
o
u
n
d
e
d
q
u
esti
o
n
s
b
u
t s
tr
u
g
g
les with
m
o
r
e
co
m
p
lex
o
r
am
b
ig
u
o
u
s
q
u
e
r
ies.
Hig
h
-
s
co
r
in
g
c
ases
co
n
f
ir
m
th
e
m
o
d
el’
s
ca
p
ab
ilit
y
to
g
en
er
ate
r
esp
o
n
s
es
clo
s
ely
m
atch
in
g
th
e
r
ef
er
en
ce
s
,
wh
ile
lo
w
-
s
co
r
i
n
g
o
u
tlier
s
r
ev
ea
l
lim
itatio
n
s
in
r
o
b
u
s
tn
ess
an
d
g
en
er
aliza
tio
n
.
Ov
er
all,
th
e
r
esu
lts
in
d
icate
p
r
o
m
is
in
g
a
n
s
wer
g
e
n
er
atio
n
a
b
ilit
y
,
th
o
u
g
h
f
u
r
th
er
im
p
r
o
v
e
m
en
ts
ar
e
n
ee
d
ed
to
r
ed
u
ce
p
er
f
o
r
m
an
ce
v
ar
ian
ce
an
d
in
cr
ea
s
e
co
n
s
is
ten
cy
ac
r
o
s
s
d
iv
er
s
e
v
is
u
al
an
d
lin
g
u
is
tic
s
ce
n
ar
io
s
.
R
esp
o
n
s
e
tim
e
co
m
p
ar
is
o
n
s
b
etwe
en
m
eth
o
d
s
ar
e
p
r
esen
ted
in
T
ab
le
4
.
T
h
e
r
es
p
o
n
s
e
tim
e
r
ep
o
r
ted
in
T
ab
le
4
r
e
v
ea
ls
an
ap
p
ar
e
n
t
p
ar
a
d
o
x
.
Alth
o
u
g
h
th
e
B
L
I
P
-
Fin
etu
n
e
m
o
d
el
(
ViT
-
B
/1
6
co
m
b
in
e
d
with
a
B
L
I
P
Dec
o
d
er
)
is
co
n
s
id
er
ab
ly
s
m
aller
in
ter
m
s
o
f
m
o
d
el
p
ar
am
eter
s
th
an
I
n
s
tr
u
ctB
L
I
P
(
ViT
-
g
/1
4
p
ai
r
ed
with
FLA
N
-
T5
-
XL
)
,
it
n
ev
er
t
h
eless
ex
h
ib
its
a
s
ig
n
if
ican
tly
h
ig
h
er
av
er
ag
e
r
esp
o
n
s
e
tim
e
(
0
.
5
5
1
6
s
co
m
p
ar
ed
to
0
.
0
9
0
7
s
)
.
T
h
is
d
is
cr
ep
a
n
cy
ca
n
n
o
t
b
e
attr
ib
u
ted
t
o
r
aw
co
m
p
u
tatio
n
al
co
m
p
le
x
ity
o
r
m
o
d
el
s
ize
alo
n
e
.
I
n
s
tead
,
th
e
p
r
im
ar
y
tech
n
ical
ca
u
s
e
lies
in
th
e
g
en
er
atio
n
a
n
d
d
ec
o
d
in
g
s
tr
ateg
y
em
p
lo
y
e
d
d
u
r
i
n
g
in
f
e
r
en
ce
.
Sp
ec
if
ically
,
th
e
B
L
I
P
-
T
5
T
P3
m
o
d
el
f
r
eq
u
en
tly
g
en
e
r
ates
lo
n
g
er
o
u
tp
u
t
s
eq
u
en
ce
s
d
u
e
to
u
n
r
eso
lv
ed
wo
r
d
r
ep
etitio
n
is
s
u
es.
As
a
r
esu
lt,
its
g
en
er
ated
r
esp
o
n
s
es
o
f
ten
r
ea
ch
th
e
p
r
e
d
ef
in
ed
m
ax
_
len
g
th
th
r
e
s
h
o
ld
(
e.
g
.
,
6
4
o
r
e
v
en
1
2
8
to
k
en
s
)
,
lead
i
n
g
to
ex
ten
d
ed
d
ec
o
d
in
g
lo
o
p
s
an
d
in
cr
ea
s
ed
laten
cy
.
I
n
ad
d
itio
n
,
th
e
s
to
p
p
i
n
g
m
ec
h
an
is
m
in
B
L
I
P
-
Fin
etu
n
e
is
less
ef
f
ec
tiv
e,
as
th
e
m
o
d
el
d
o
es
n
o
t
c
o
n
s
is
ten
tly
p
r
ed
ict
t
h
e
en
d
-
of
-
s
eq
u
en
ce
(
<E
OS>)
to
k
en
at
an
ea
r
l
y
s
tag
e.
I
n
co
n
tr
ast,
I
n
s
tr
u
ctB
L
I
P
an
d
B
L
I
P
-
2
t
y
p
ically
p
r
o
d
u
c
e
m
u
ch
s
h
o
r
ter
a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
42
,
No
.
3
,
J
u
n
e
20
2
6
:
7
4
2
-
75
2
750
m
o
r
e
c
o
n
cise
r
esp
o
n
s
es,
o
f
ten
co
n
tain
in
g
f
ewe
r
th
an
1
0
t
o
k
en
s
.
T
h
ese
m
o
d
els
ten
d
to
p
r
e
d
ict
th
e
<E
OS>
to
k
en
v
er
y
ea
r
l
y
in
th
e
d
ec
o
d
in
g
p
r
o
ce
s
s
,
allo
win
g
g
en
er
ati
o
n
to
ter
m
in
ate
p
r
o
m
p
tly
.
C
o
n
s
eq
u
en
tly
,
d
esp
ite
th
eir
lar
g
er
p
ar
am
eter
s
izes,
th
eir
in
f
er
en
ce
tim
e
r
em
ain
s
s
u
b
s
tan
tially
lo
wer
d
u
e
to
r
ed
u
ce
d
to
k
en
-
by
-
to
k
en
d
ec
o
d
in
g
o
v
e
r
h
ea
d
.
T
ab
le
4
.
R
esp
o
n
s
e
tim
e
r
esu
lts
b
etwe
en
m
eth
o
d
s
M
e
t
h
o
d
B
a
c
k
b
o
n
e
A
V
G
R
e
s
p
o
n
se
t
i
me
(
s)
B
LI
P
-
2
V
i
T
-
g
/
1
4
+
F
LA
N
-
T5
-
XL
0
.
2
1
8
9
I
n
st
r
u
c
t
B
LI
P
V
i
T
-
g
/
1
4
+
F
LA
N
-
T5
-
XL
0
.
0
9
0
7
B
LI
P
-
T5
TP3
(
o
u
r
s)
V
i
T
-
B
/
1
6
+
B
LI
P
D
e
c
o
d
e
r
0
.
5
5
1
6
D
Ei
T
+
B
e
r
t
D
Ei
T
-
B
/
1
6
+
B
ER
T
-
b
a
se
0
.
0
3
5
1
B
Ei
T
+
G
P
T2
V
i
T
-
B
/
1
6
+
g
p
t
2
0
.
1
7
8
1
W
e
an
aly
ze
d
all
in
co
r
r
ec
t
p
r
e
d
ictio
n
s
o
f
th
e
test
s
et
an
d
s
aw
s
ev
er
al
in
h
er
e
n
t
lim
itatio
n
s
o
f
th
e
B
L
I
P
-
b
ased
VQA
m
o
d
el.
First,
th
e
m
o
d
el
s
tr
u
g
g
les
with
q
u
esti
o
n
s
th
at
r
eq
u
ir
e
b
ac
k
g
r
o
u
n
d
o
r
wo
r
ld
k
n
o
wled
g
e
b
ey
o
n
d
wh
at
is
ex
p
licitly
v
is
ib
le
in
th
e
im
ag
e,
s
u
ch
as
id
e
n
tify
in
g
th
e
n
am
e
o
f
th
e
“Ok
l
ah
o
m
a
Un
i
v
er
s
ity
”
f
o
o
tb
all
team
.
T
h
is
lim
itatio
n
ar
is
es
b
ec
au
s
e
th
e
cu
r
r
e
n
t
s
y
s
tem
d
o
es
n
o
t
i
n
co
r
p
o
r
ate
e
x
ter
n
al
k
n
o
wled
g
e
r
etr
iev
al
m
ec
h
an
is
m
s
o
r
tex
t
r
ec
o
g
n
itio
n
(
OC
R
)
m
o
d
u
les
th
at
co
u
ld
p
r
o
v
id
e
ad
d
itio
n
al
s
e
m
an
tic
in
f
o
r
m
atio
n
.
Seco
n
d
,
th
e
m
o
d
e
l
d
e
m
o
n
s
tr
ates
lim
ited
ca
p
ab
ilit
y
in
q
u
an
titativ
e
r
ea
s
o
n
in
g
an
d
f
in
e
-
g
r
ain
ed
s
em
an
tic
u
n
d
er
s
tan
d
i
n
g
,
as
ev
id
e
n
ce
d
b
y
er
r
o
r
s
in
c
o
u
n
tin
g
th
e
n
u
m
b
er
o
f
p
e
o
p
le
in
a
n
im
ag
e
o
r
r
e
co
g
n
izin
g
c
o
m
p
lex
em
o
tio
n
al
s
tates
s
u
ch
as
“e
x
cited
.
”
T
h
ese
task
s
r
eq
u
ir
e
n
o
t
o
n
ly
o
b
ject
d
etec
tio
n
b
u
t
also
co
n
tex
tu
al
in
ter
p
r
etatio
n
o
f
f
ac
ial
ex
p
r
ess
io
n
s
,
b
o
d
y
lan
g
u
ag
e,
an
d
g
r
o
u
p
d
y
n
a
m
ics,
wh
ich
a
r
e
n
o
t
ex
p
licitly
m
o
d
eled
.
Fin
ally
,
th
e
s
y
s
tem
is
p
r
o
n
e
t
o
o
b
ject
an
d
ac
tio
n
co
n
f
u
s
io
n
,
f
o
r
i
n
s
tan
ce
m
is
tak
in
g
a
“
ca
n
n
o
n
”
f
o
r
a
“tir
e
s
win
g
,
”
in
d
icatin
g
wea
k
n
ess
es
in
r
ec
o
g
n
izin
g
s
p
ec
ialized
o
b
jects
an
d
in
f
er
r
in
g
co
llec
tiv
e
o
r
ev
en
t
-
lev
el
ac
tio
n
s
.
Ov
er
all,
th
ese
er
r
o
r
s
h
ig
h
lig
h
t
th
at
th
e
m
o
d
el
r
e
lies
p
r
im
ar
ily
o
n
v
is
u
al
f
ea
tu
r
es
an
d
q
u
esti
o
n
p
r
o
m
p
ts
,
with
o
u
t
s
u
p
p
o
r
t
f
r
o
m
ex
ter
n
al
k
n
o
wled
g
e
s
o
u
r
ce
s
o
r
s
p
ec
ialized
r
ea
s
o
n
in
g
m
o
d
u
les,
wh
ich
lim
its
its
r
o
b
u
s
tn
ess
in
co
m
p
le
x
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
.
4.
CO
NCLU
SI
O
N
T
h
is
s
tu
d
y
m
ak
es
s
ev
er
al
m
ea
s
u
r
ab
le
co
n
tr
ib
u
tio
n
s
to
VQA
r
esear
ch
b
y
im
p
r
o
v
in
g
i
n
ter
o
p
er
a
b
ilit
y
an
d
au
to
m
atic
an
s
wer
g
en
er
a
tio
n
th
r
o
u
g
h
th
e
i
n
teg
r
atio
n
o
f
a
B
L
I
P
-
T
5
VQA
m
o
d
el
with
an
au
to
m
ated
q
u
esti
o
n
g
en
er
atio
n
p
ip
elin
e.
A
n
ew
VQA
d
ataset
was
co
n
s
tr
u
cted
f
r
o
m
th
e
Fli
ck
r
8
k
d
a
taset,
wh
er
e
im
ag
e
ca
p
tio
n
s
wer
e
au
to
m
atica
lly
tr
an
s
f
o
r
m
ed
in
t
o
q
u
esti
o
n
–
an
s
wer
p
air
s
u
s
in
g
th
e
T
5
T
P3
m
o
d
el,
en
ab
lin
g
sc
alab
le
d
ata
cr
ea
tio
n
with
o
u
t
m
an
u
al
an
n
o
tatio
n
.
Fin
e
-
t
u
n
in
g
B
L
I
P
-
T
5
VQA
o
n
th
i
s
d
ataset
p
r
o
d
u
ce
d
co
n
s
is
ten
t
im
p
r
o
v
e
m
en
ts
o
v
e
r
b
aselin
e
m
o
d
els
s
u
ch
as
B
L
I
P
-
2
,
I
n
s
tr
u
ctB
L
I
P,
DE
iT
+
B
er
t,
an
d
B
E
iT
+
GPT2
ac
co
r
d
in
g
to
B
L
E
U
-
1
to
B
L
E
U
-
4
a
n
d
R
OUGE
-
L
m
etr
ics.
E
x
p
e
r
im
en
tal
r
esu
lt
s
s
h
o
w
p
ar
ticu
lar
ly
s
tr
o
n
g
g
ai
n
s
in
B
L
E
U
-
1
an
d
B
L
E
U
-
2
,
r
e
f
lectin
g
ac
c
u
r
ate
r
ec
o
g
n
itio
n
o
f
k
ey
v
is
u
al
c
o
n
ce
p
ts
an
d
s
h
o
r
t
an
s
wer
s
,
wh
ile
lo
wer
B
L
E
U
-
3
a
n
d
B
L
E
U
-
4
s
co
r
es
in
d
i
ca
te
o
n
g
o
in
g
ch
allen
g
es
in
g
en
er
atin
g
lo
n
g
er
,
co
m
p
o
s
itio
n
al
r
esp
o
n
s
es.
R
OUGE
-
L
r
esu
lts
d
em
o
n
s
tr
ate
m
o
d
er
ate
s
tr
u
ctu
r
al
s
im
ilar
ity
to
r
ef
er
en
ce
an
s
wer
s
,
th
o
u
g
h
with
s
o
m
e
v
ar
ia
n
ce
ac
r
o
s
s
s
am
p
les.
Desp
ite
th
ese
s
tr
en
g
th
s
,
th
e
ap
p
r
o
ac
h
s
till
h
as
lim
itatio
n
s
.
Per
f
o
r
m
a
n
ce
m
ay
d
ec
lin
e
o
n
co
m
p
lex
r
ea
s
o
n
in
g
task
s
r
eq
u
i
r
in
g
e
x
ten
s
iv
e
e
x
ter
n
al
k
n
o
wled
g
e
o
r
m
u
lti
-
s
tep
i
n
f
er
e
n
ce
,
a
n
d
th
e
clar
if
icatio
n
m
ec
h
an
is
m
r
elies
o
n
p
r
ed
e
f
in
ed
p
r
o
m
p
tin
g
h
eu
r
is
tics
r
ath
er
th
an
ad
a
p
tiv
e
d
ialo
g
u
e
p
o
licies.
I
n
ad
d
itio
n
,
ev
alu
atio
n
s
wer
e
co
n
d
u
cte
d
m
ain
ly
o
n
b
en
c
h
m
ar
k
d
atasets
an
d
lim
ited
r
ea
l
-
wo
r
ld
s
ce
n
ar
i
o
s
,
wh
ich
m
ay
n
o
t
f
u
lly
ca
p
tu
r
e
d
iv
er
s
e
u
s
er
b
e
h
av
io
r
.
Fu
tu
r
e
r
esear
ch
s
h
o
u
ld
in
v
esti
g
ate
ad
ap
tiv
e
p
r
o
m
p
t
le
ar
n
in
g
,
p
ar
am
eter
-
ef
f
icien
t
f
in
e
-
tu
n
in
g
,
m
u
lti
-
tu
r
n
d
ialo
g
u
e
m
o
d
elin
g
,
an
d
la
r
g
e
-
s
ca
le
u
s
er
-
in
-
th
e
-
lo
o
p
ev
a
lu
atio
n
s
to
f
u
r
t
h
er
im
p
r
o
v
e
r
o
b
u
s
tn
ess
,
g
en
er
aliz
atio
n
,
an
d
p
r
ac
tical
ap
p
licab
ilit
y
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
I
mp
r
o
ve
d
in
tera
ctivity
a
n
d
a
u
t
o
ma
ted
r
esp
o
n
s
e
fo
r
visu
a
l q
u
esti
o
n
a
n
s
w
erin
g
…
(
N
g
u
ye
n
Ha
Ma
n
h
K
h
a
n
g
)
751
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Ng
u
y
en
Ha
Ma
n
h
Kh
an
g
✓
✓
✓
✓
✓
✓
✓
✓
Ng
u
y
en
T
u
an
An
h
✓
✓
✓
✓
Ng
u
y
en
Min
h
Ho
an
g
✓
✓
✓
✓
B
u
i T
h
an
h
Hu
n
g
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
in
d
in
g
s
o
f
th
is
s
tu
d
y
ar
e
o
p
e
n
ly
av
ailab
le
in
[
U
n
iv
er
s
ity
o
f
I
llin
o
is
Ur
b
an
a
-
C
h
am
p
aig
n
]
at
h
ttp
s
://fo
r
m
s
.
illi
n
o
is
.
ed
u
/s
ec
/1
7
1
3
3
9
8
,
r
ef
er
e
n
ce
n
u
m
b
er
[
2
2
]
.
RE
F
E
R
E
NC
E
S
[
1
]
B
.
T.
H
u
n
g
a
n
d
H
.
V
.
H
.
D
u
y
,
“
E
x
V
Q
A
:
a
n
o
v
e
l
st
a
c
k
e
d
a
t
t
e
n
t
i
o
n
n
e
t
w
o
r
k
s
w
i
t
h
e
x
t
e
n
d
e
d
l
o
n
g
sh
o
r
t
-
t
e
r
m
m
e
mo
r
y
m
o
d
e
l
f
o
r
v
i
s
u
a
l
q
u
e
st
i
o
n
a
n
sw
e
r
i
n
g
,
”
C
o
m
p
u
t
e
rs
a
n
d
El
e
c
t
ri
c
a
l
E
n
g
i
n
e
e
ri
n
g
,
v
o
l
.
1
2
6
,
p
.
1
1
0
4
3
9
,
A
u
g
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
o
mp
e
l
e
c
e
n
g
.
2
0
2
5
.
1
1
0
4
3
9
.
[
2
]
S.
Lu
,
M
.
L
i
u
,
L.
Y
i
n
,
Z
.
Y
i
n
,
X
.
Li
u
,
a
n
d
W
.
Z
h
e
n
g
,
“
T
h
e
mu
l
t
i
-
mo
d
a
l
f
u
s
i
o
n
i
n
v
i
s
u
a
l
q
u
e
st
i
o
n
a
n
sw
e
r
i
n
g
:
a
r
e
v
i
e
w
o
f
a
t
t
e
n
t
i
o
n
mec
h
a
n
i
sms
,
”
Pe
e
rJ
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
9
,
p
.
e
1
4
0
0
,
M
a
y
2
0
2
3
,
d
o
i
:
1
0
.
7
7
1
7
/
p
e
e
r
j
-
c
s.
1
4
0
0
.
[
3
]
S
.
Y
a
n
g
,
C
.
H
a
n
,
S
.
L
u
o
,
a
n
d
E.
H
o
v
y
,
“
M
A
G
I
C
-
V
Q
A
:
M
u
l
t
i
m
o
d
a
l
a
n
d
g
r
o
u
n
d
e
d
i
n
f
e
r
e
n
c
e
w
i
t
h
c
o
mm
o
n
s
e
n
se
k
n
o
w
l
e
d
g
e
f
o
r
v
i
s
u
a
l
q
u
e
st
i
o
n
a
n
sw
e
r
i
n
g
,
”
i
n
Fi
n
d
i
n
g
s
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s:
A
C
L
2
0
2
5
,
S
t
r
o
u
d
sb
u
r
g
,
P
A
,
U
S
A
:
A
sso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
2
0
2
5
,
p
p
.
1
6
9
6
7
–
1
6
9
8
6
.
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
5
.
f
i
n
d
i
n
g
s
-
a
c
l
.
8
7
2
.
[
4
]
A
.
P
a
n
d
e
y
,
D
.
B
o
d
o
,
A
.
P
h
u
k
a
n
,
a
n
d
A
.
E
k
b
a
l
,
“
T
h
e
q
u
e
st
f
o
r
v
i
s
u
a
l
u
n
d
e
r
st
a
n
d
i
n
g
:
a
j
o
u
r
n
e
y
t
h
r
o
u
g
h
t
h
e
e
v
o
l
u
t
i
o
n
o
f
v
i
su
a
l
q
u
e
st
i
o
n
a
n
sw
e
r
i
n
g
.
”
Ja
n
.
1
3
,
2
0
2
5
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
2
5
0
1
.
0
7
1
0
9
[
5
]
H
.
J
.
S
i
n
g
h
,
G
.
B
a
t
h
l
a
,
M
.
M
e
h
t
a
,
G
.
C
h
h
a
b
r
a
,
a
n
d
P
.
S
i
n
g
h
,
“
V
i
s
u
a
l
q
u
e
s
t
i
o
n
s
a
n
s
w
e
r
i
n
g
d
e
v
e
l
o
p
m
e
n
t
s
,
a
p
p
l
i
c
a
t
i
o
n
s
,
d
a
t
a
s
e
t
s
a
n
d
o
p
p
o
r
t
u
n
i
t
i
e
s
:
A
s
t
a
t
e
-
of
-
t
h
e
-
a
r
t
s
u
r
v
e
y
,
”
i
n
2
n
d
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
S
u
s
t
a
i
n
a
b
l
e
C
o
m
p
u
t
i
n
g
a
n
d
D
a
t
a
C
o
m
m
u
n
i
c
a
t
i
o
n
S
y
s
t
e
m
s
,
I
C
S
C
D
S
2
0
2
3
-
P
r
o
c
e
e
d
i
n
g
s
,
I
E
E
E
,
M
a
r
.
2
0
2
3
,
p
p
.
7
7
8
–
7
8
5
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
S
C
D
S
5
6
5
8
0
.
2
0
2
3
.
1
0
1
0
4
8
7
0
.
[
6
]
B
.
T
.
H
u
n
g
,
“
C
o
n
t
e
n
t
-
b
a
se
d
i
m
a
g
e
r
e
t
r
i
e
v
a
l
u
si
n
g
m
u
l
t
i
-
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
s,”
i
n
L
e
c
t
u
re
N
o
t
e
s
i
n
N
e
t
w
o
rks
a
n
d
S
y
s
t
e
m
s
,
v
o
l
.
4
4
5
,
2
0
2
3
,
p
p
.
3
4
7
–
3
5
7
.
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
981
-
19
-
1
4
1
2
-
6
_
2
9
.
[
7
]
B
.
T.
H
u
n
g
,
“
Li
n
k
p
r
e
d
i
c
t
i
o
n
i
n
p
a
p
e
r
c
i
t
a
t
i
o
n
n
e
t
w
o
r
k
b
a
s
e
d
o
n
d
e
e
p
g
r
a
p
h
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
L
e
c
t
u
re
N
o
t
e
s
o
n
D
a
t
a
E
n
g
i
n
e
e
ri
n
g
a
n
d
C
o
m
m
u
n
i
c
a
t
i
o
n
s Te
c
h
n
o
l
o
g
i
e
s
,
v
o
l
.
1
1
7
,
2
0
2
2
,
p
p
.
8
9
7
–
9
0
7
.
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
9
8
1
-
19
-
0898
-
9
_
6
7
.
[
8
]
B
.
T
.
H
u
n
g
a
n
d
V
.
Q
.
H
u
y
,
“
M
TFI
C
:
e
n
h
a
n
c
e
d
f
a
sh
i
o
n
i
ma
g
e
c
a
p
t
i
o
n
i
n
g
v
i
a
mu
l
t
i
-
t
r
a
n
sf
o
r
mer
a
r
c
h
i
t
e
c
t
u
r
e
w
i
t
h
c
o
n
t
r
a
s
t
i
v
e
a
n
d
b
i
d
i
r
e
c
t
i
o
n
a
l
e
n
c
o
d
i
n
g
s,
”
V
i
s
u
a
l
C
o
m
p
u
t
e
r
,
v
o
l
.
4
1
,
n
o
.
1
3
,
p
p
.
1
0
8
4
1
–
1
0
8
5
5
,
O
c
t
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
3
7
1
-
0
2
5
-
0
4
0
7
2
-
8.
[
9
]
B
.
T
.
H
u
n
g
,
N
.
V
.
P
.
N
h
a
n
,
a
n
d
N
.
T.
S
y
,
“
D
C
A
R
ES:
d
e
e
p
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
w
i
t
h
n
e
u
r
a
l
-
b
a
se
d
o
p
t
i
m
i
z
a
t
i
o
n
f
o
r
i
ma
g
e
-
b
a
se
d
p
r
o
d
u
c
t
r
e
c
o
mm
e
n
d
e
r
s
y
st
e
m,”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
4
,
n
o
.
3
0
,
p
p
.
3
6
6
9
3
–
3
6
7
2
3
,
F
e
b
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
0
4
2
-
025
-
2
0
6
5
5
-
y.
[
1
0
]
S
.
C
h
o
w
d
h
u
r
y
a
n
d
B
.
S
o
n
i
,
“
R
-
V
Q
A
:
A
r
o
b
u
st
v
i
s
u
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
m
o
d
e
l
,
”
K
n
o
w
l
e
d
g
e
-
B
a
s
e
d
S
y
st
e
m
s
,
v
o
l
.
3
0
9
,
p
.
1
1
2
8
2
7
,
Ja
n
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
k
n
o
sy
s.
2
0
2
4
.
1
1
2
8
2
7
.
[
1
1
]
N
.
D
.
H
u
y
n
h
,
M
.
R
.
B
o
u
a
d
j
e
n
e
k
,
S
.
A
r
y
a
l
,
I
.
R
a
z
z
a
k
,
a
n
d
H
.
H
a
c
i
d
,
“
V
i
su
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
:
f
r
o
m
e
a
r
l
y
d
e
v
e
l
o
p
me
n
t
s
t
o
r
e
c
e
n
t
a
d
v
a
n
c
e
s
--
a
s
u
r
v
e
y
.
”
Ja
n
.
1
1
,
2
0
2
5
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
2
5
0
1
.
0
3
9
3
9
[
1
2
]
M
.
Y
a
ma
d
a
,
V
.
D
’
a
m
a
r
i
o
,
K
.
Ta
k
e
m
o
t
o
,
X
.
B
o
i
x
,
a
n
d
T
.
S
a
sa
k
i
,
“
Tr
a
n
sf
o
r
mer
m
o
d
u
l
e
n
e
t
w
o
r
k
s
f
o
r
sy
s
t
e
m
a
t
i
c
g
e
n
e
r
a
l
i
z
a
t
i
o
n
i
n
v
i
s
u
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
,
”
I
E
EE
T
ra
n
s
a
c
t
i
o
n
s
o
n
Pa
t
t
e
rn
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
.
4
6
,
n
o
.
1
2
,
p
p
.
1
0
0
9
6
–
1
0
1
0
5
,
M
a
r
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
2
4
.
3
4
3
8
8
8
7
.
[
1
3
]
S
.
A
n
t
o
l
e
t
a
l
.
,
“
V
q
a
:
v
i
su
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
E
EE
i
n
t
e
r
n
a
t
i
o
n
a
l
c
o
n
f
e
re
n
c
e
o
n
c
o
m
p
u
t
e
r
v
i
s
i
o
n
,
2
0
1
5
,
p
p
.
2
4
2
5
–
2
4
3
3
.
[
1
4
]
Q
.
W
u
,
D
.
Te
n
e
y
,
P
.
W
a
n
g
,
C
.
S
h
e
n
,
A
.
D
i
c
k
,
a
n
d
A
.
v
a
n
d
e
n
H
e
n
g
e
l
,
“
V
i
su
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
:
a
s
u
r
v
e
y
o
f
m
e
t
h
o
d
s
a
n
d
d
a
t
a
se
t
s,
”
C
o
m
p
u
t
e
r
Vi
s
i
o
n
a
n
d
I
m
a
g
e
U
n
d
e
rs
t
a
n
d
i
n
g
,
v
o
l
.
1
6
3
,
p
p
.
2
1
–
4
0
,
2
0
1
7
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
v
i
u
.
2
0
1
7
.
0
5
.
0
0
1
.
[
1
5
]
R
.
K
a
b
i
r
,
N
.
H
a
q
u
e
,
M
.
S
.
I
sl
a
m,
a
n
d
M
a
r
i
u
m
-
E
-
Ja
n
n
a
t
,
“
A
c
o
m
p
r
e
h
e
n
s
i
v
e
s
u
r
v
e
y
o
n
v
i
s
u
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
d
a
t
a
set
s
a
n
d
a
l
g
o
r
i
t
h
ms
.
”
N
o
v
.
1
7
,
2
0
2
4
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
2
4
1
1
.
1
1
1
5
0
[
1
6
]
J.
L
i
,
D
.
L
i
,
C
.
X
i
o
n
g
,
a
n
d
S
.
H
o
i
,
“
B
LI
P
:
b
o
o
t
st
r
a
p
p
i
n
g
l
a
n
g
u
a
g
e
-
i
m
a
g
e
p
r
e
-
t
r
a
i
n
i
n
g
f
o
r
u
n
i
f
i
e
d
v
i
s
i
o
n
-
l
a
n
g
u
a
g
e
u
n
d
e
r
st
a
n
d
i
n
g
a
n
d
g
e
n
e
r
a
t
i
o
n
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
Ma
c
h
i
n
e
L
e
a
rn
i
n
g
Re
s
e
a
r
c
h
,
v
o
l
.
1
6
2
,
p
p
.
1
2
8
8
8
–
1
2
9
0
0
,
2
0
2
2
.
[
1
7
]
C
.
Zh
a
n
g
,
H
.
Zh
a
n
g
,
Y
.
S
u
n
,
a
n
d
J.
W
a
n
g
,
“
D
o
w
n
s
t
r
e
a
m
t
r
a
n
sf
o
r
mer
g
e
n
e
r
a
t
i
o
n
o
f
q
u
e
s
t
i
o
n
-
a
n
sw
e
r
p
a
i
r
s
w
i
t
h
p
r
e
p
r
o
c
e
ssi
n
g
a
n
d
p
o
s
t
p
r
o
c
e
ssi
n
g
p
i
p
e
l
i
n
e
s,”
i
n
D
o
c
E
n
g
2
0
2
2
-
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
2
2
A
C
M
S
y
m
p
o
si
u
m
o
n
D
o
c
u
m
e
n
t
E
n
g
i
n
e
e
ri
n
g
,
2
0
2
2
.
d
o
i
:
1
0
.
1
1
4
5
/
3
5
5
8
1
0
0
.
3
5
6
3
8
4
6
.
[
1
8
]
K
.
Z
h
u
e
t
a
l
.
,
“
P
r
o
m
p
t
R
o
b
u
st
:
To
w
a
r
d
s
e
v
a
l
u
a
t
i
n
g
t
h
e
r
o
b
u
st
n
e
ss
o
f
l
a
r
g
e
l
a
n
g
u
a
g
e
m
o
d
e
l
s
o
n
a
d
v
e
r
sari
a
l
p
r
o
m
p
t
s
,
”
i
n
L
A
MP
S
2
0
2
4
-
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
1
st
A
C
M
W
o
rks
h
o
p
o
n
L
a
r
g
e
AI
S
y
s
t
e
m
s
a
n
d
Mo
d
e
l
s
w
i
t
h
Pr
i
v
a
c
y
a
n
d
S
a
f
e
t
y
An
a
l
y
s
i
s
,
N
e
w
Y
o
r
k
,
N
Y
,
U
S
A
:
A
C
M
,
N
o
v
.
2
0
2
4
,
p
p
.
5
7
–
6
8
.
d
o
i
:
1
0
.
1
1
4
5
/
3
6
8
9
2
1
7
.
3
6
9
0
6
2
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.