I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
,
p
p
.
941
~
952
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
1
.
p
p
9
4
1
-
9
5
2
941
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Adv
ersa
ria
l ex
a
m
ples in A
ra
bic lan
g
ua
g
e
Sa
f
a
e
L
a
a
t
y
a
o
ui,
M
o
ha
mm
e
d Sa
ber
S
martI
C
T
La
b
o
r
a
t
o
r
y
,
EN
S
A
O
,
M
o
h
a
mm
e
d
F
i
r
s
t
U
n
i
v
e
r
s
i
t
y
,
O
u
j
d
a
,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
1
6
,
2
0
2
4
R
ev
is
ed
No
v
1
4
,
2
0
2
5
Acc
ep
ted
J
an
1
0
,
2
0
2
6
Ad
v
e
rsa
rial
a
tt
a
c
k
s
h
a
v
e
a
g
re
a
t
p
o
p
u
larit
y
in
th
e
a
rti
ficia
l
in
tell
ig
e
n
c
e
(AI)
d
o
m
a
in
.
In
th
e
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
(NLP
)
field
,
v
a
ri
o
u
s
t
e
c
h
n
iq
u
e
s
h
a
v
e
b
e
e
n
u
se
d
t
o
e
v
a
lu
a
te t
h
e
v
u
ln
e
ra
b
il
it
y
o
f
d
e
e
p
lea
rn
in
g
(DL)
m
o
d
e
ls.
It
is
o
b
se
rv
e
d
t
h
a
t
w
h
il
e
m
o
st
st
u
d
ies
f
o
c
u
se
d
o
n
g
e
n
e
ra
ti
n
g
a
d
v
e
rsa
rial
e
x
a
m
p
les
in
En
g
li
s
h
lan
g
u
a
g
e
,
Ar
a
b
ic
a
d
v
e
rsa
rial
a
tt
a
c
k
s
h
a
v
e
re
c
e
iv
e
d
li
tt
le
a
tt
e
n
ti
o
n
.
T
h
is
p
a
p
e
r
p
re
se
n
t
s
a
two
-
ste
p
m
e
th
o
d
to
c
re
a
te
a
d
v
e
rsa
rial
e
x
a
m
p
les
in
Ara
b
ic
lan
g
u
a
g
e
:
fir
st,
th
e
m
o
st
imp
o
rtan
t
wo
r
d
s
a
re
id
e
n
ti
fie
d
.
Th
e
n
,
t
h
e
p
ro
p
o
se
d
tra
n
sfo
rm
a
ti
o
n
a
lg
o
rit
h
m
is
a
p
p
li
e
d
.
On
ly
s
m
a
ll
a
n
d
imp
e
rc
e
p
ti
b
le
m
a
n
ip
u
lati
o
n
s
b
a
se
d
o
n
c
o
m
m
o
n
m
istak
e
s
in
Ara
b
ic
writi
n
g
m
islea
d
th
e
p
o
p
u
lar
p
re
-
trai
n
e
d
lan
g
u
a
g
e
m
o
d
e
l
(P
LM
)
b
i
d
irec
ti
o
n
a
l
e
n
c
o
d
e
r
re
p
re
se
n
tatio
n
s
fr
o
m
tran
sf
o
rm
e
rs
(BERT
)
re
train
e
d
o
n
t
h
e
b
o
o
k
re
v
iew
s
in
Ara
b
ic
d
a
tas
e
t
(BRAD
)
o
n
th
e
se
n
ti
m
e
n
t
a
n
a
ly
sis
(S
A)
tas
k
a
n
d
d
e
c
re
a
s
e
it
s
p
e
rfo
rm
a
n
c
e
:
th
e
c
las
sifica
ti
o
n
a
c
c
u
ra
c
y
wa
s
re
d
u
c
e
d
b
y
a
n
a
v
e
ra
g
e
o
f
3
.
4
4
%
.
T
h
is
d
r
o
p
in
a
c
c
u
ra
c
y
sh
o
ws
t
h
a
t
t
h
e
m
o
d
e
l
wa
s
su
c
c
e
ss
fu
ll
y
a
tt
a
c
k
e
d
.
K
ey
w
o
r
d
s
:
Ad
v
er
s
ar
ial
attac
k
B
E
R
T
L
o
w
-
r
eso
u
r
ce
la
n
g
u
a
g
e
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
Pre
-
tr
ain
ed
lan
g
u
ag
e
m
o
d
el
Sen
tim
en
t a
n
aly
s
is
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
:
Saf
ae
L
aa
ty
ao
u
i
Sm
ar
tI
C
T
L
ab
o
r
ato
r
y
,
E
NSAO,
Mo
h
am
m
ed
First Un
iv
er
s
ity
Ou
jd
a
6
0
0
0
0
,
Mo
r
o
cc
o
E
m
ail:
s
.
laaty
ao
u
i@
u
m
p
.
ac
.
m
a
1.
I
NT
RO
D
UCT
I
O
N
Dee
p
lear
n
in
g
(
DL
)
m
o
d
els
ar
e
v
u
ln
er
ab
le
to
ad
v
er
s
ar
ial
ex
am
p
les
;
p
er
tu
r
b
e
d
in
p
u
ts
th
a
t
m
ay
b
e
im
p
er
ce
p
tib
le
t
o
th
e
h
u
m
an
ey
e
co
u
ld
f
o
r
ce
t
h
e
s
y
s
tem
to
p
r
o
d
u
ce
in
co
r
r
ec
t
o
u
tp
u
ts
[
1
]
–
[
3
]
.
Mo
s
t o
f
t
h
e
ea
r
l
y
s
tu
d
ies
f
o
cu
s
o
n
im
ag
e
in
p
u
t
s
[
4
]
–
[
7
]
.
As
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P
)
h
as
ad
v
an
ce
d
s
ig
n
if
ican
tly
,
s
tu
d
ies ar
e
also
f
o
cu
s
ed
o
n
ev
alu
atin
g
th
e
r
o
b
u
s
tn
ess
o
f
NL
P m
o
d
els u
s
in
g
ad
v
e
r
s
ar
ial
attac
k
s
[
8
]
–
[
1
2
]
.
Mo
s
t
wo
r
k
s
in
th
e
NL
P
f
ield
f
o
cu
s
o
n
co
n
s
tr
u
ctin
g
ad
v
e
r
s
ar
ial
ex
am
p
les
in
E
n
g
lis
h
.
T
ex
t
Gu
is
e
[
1
3
]
is
an
ad
a
p
tiv
e
b
lack
b
o
x
(
B
B
)
m
eth
o
d
th
at
u
s
es
p
er
tu
r
b
atio
n
s
at
b
o
th
wo
r
d
an
d
s
en
t
en
ce
g
r
an
u
lar
ity
to
p
r
eser
v
e
s
em
an
tics
an
d
f
o
o
l
class
if
ier
s
.
Dee
p
W
o
r
d
B
u
g
[
1
4
]
d
ef
in
es
s
co
r
in
g
s
tr
ateg
ies
to
id
en
tify
th
e
to
k
en
s
th
at,
if
m
o
d
if
ied
,
will
ch
a
n
g
e
t
h
e
p
r
e
d
ictio
n
o
f
th
e
class
if
ier
i
n
B
B
s
ettin
g
s
.
Qi
et
a
l.
[
1
5
]
p
r
o
p
o
s
es
a
g
r
a
d
ien
t
-
g
u
id
ed
,
wo
r
d
-
lev
el
m
eth
o
d
f
o
r
g
en
er
atin
g
a
d
v
er
s
ar
ial
e
x
am
p
les
th
at
f
o
o
l
tex
t
class
if
icatio
n
m
o
d
els.
T
o
ev
alu
ate
an
ex
tr
em
e
m
u
ltil
ab
e
l
tex
t
class
if
icatio
n
(
XM
T
C
)
s
y
s
tem
,
Qar
ae
i
an
d
B
ab
b
ar
[
1
6
]
g
en
e
r
ates
wo
r
d
-
lev
el
ad
v
e
r
s
ar
ial
ex
am
p
les
b
y
m
ask
in
g
a
wo
r
d
a
n
d
t
h
en
r
ep
l
ac
in
g
it
with
a
tr
an
s
f
o
r
m
er
-
p
r
ed
icted
alter
n
ativ
e.
W
an
g
et
a
l.
[
1
7
]
p
r
o
p
o
s
es
a
wo
r
d
-
lev
el
attac
k
th
at
em
p
lo
y
s
a
p
ar
allel
p
ar
ticle
s
war
m
o
p
tim
izatio
n
(
PS
O)
alg
o
r
ith
m
to
cr
af
t
a
d
v
er
s
ar
ial
tex
t
ex
am
p
les.
T
h
e
m
eth
o
d
[
1
8
]
i
n
tr
o
d
u
ce
s
a
B
B
attac
k
o
n
n
e
u
r
al
r
a
n
k
in
g
m
o
d
els.
I
t
u
s
es
a
s
u
r
r
o
g
ate
m
o
d
el
an
d
p
er
f
o
r
m
s
wo
r
d
-
s
u
b
s
t
itu
tio
n
p
e
r
tu
r
b
atio
n
s
to
p
r
o
m
o
te
th
e
r
a
n
k
in
g
o
f
a
tar
g
et
d
o
cu
m
e
n
t.
Kh
an
et
a
l.
[
1
9
]
p
r
o
p
o
s
e
a
ch
ar
ac
ter
-
lev
el
attac
k
th
at
g
en
er
ates
ad
v
er
s
ar
i
al
tex
t
g
u
id
ed
b
y
a
b
ea
m
-
s
ea
r
ch
.
I
t su
cc
ess
f
u
lly
f
o
o
ls
tex
t c
lass
if
ier
s
o
n
s
o
cial
m
ed
ia
s
ty
le
tex
ts
.
Fu
r
s
o
v
et
a
l.
[
2
0
]
p
r
o
p
o
s
es a
B
B
s
en
ten
ce
-
lev
el
attac
k
f
in
e
-
tu
n
in
g
a
p
r
e
-
tr
ai
n
ed
lan
g
u
ag
e
m
o
d
el
(
PLM
)
to
g
e
n
er
ate
ad
v
er
s
ar
ial
ex
am
p
les
.
So
m
e
s
tu
d
ies
ex
am
in
ed
ad
v
er
s
ar
ial
ex
am
p
les
g
en
e
r
atio
n
in
h
ig
h
-
r
eso
u
r
ce
lan
g
u
ag
es
(
HR
L
)
o
th
er
th
an
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
941
-
9
5
2
942
E
n
g
lis
h
s
u
ch
as
C
h
in
ese
[
2
1
]
–
[
2
3
]
,
R
u
s
s
ian
[
2
4
]
,
an
d
J
ap
a
n
ese
[
2
5
]
;
o
th
er
s
p
r
o
p
o
s
e
atta
ck
s
in
lo
w
-
r
eso
u
r
ce
lan
g
u
ag
es
(
L
R
L
)
lik
e
B
asq
u
e
l
an
g
u
ag
e
[
2
6
]
.
I
n
Ar
ab
ic
lan
g
u
ag
e,
o
n
e
o
f
th
e
L
R
L
s
,
o
n
ly
a
f
ew
wo
r
k
s
h
av
e
ex
am
in
e
d
ad
v
e
r
s
ar
ial
ex
am
p
les
.
T
h
er
ef
o
r
e,
th
is
s
tu
d
y
aim
s
to
in
v
esti
g
ate
th
e
ad
v
er
s
ar
ial
r
o
b
u
s
tn
ess
o
f
a
PLM
m
o
d
el
u
s
ed
f
o
r
b
in
ar
y
s
en
tim
en
t
an
aly
s
is
(
SA
)
task
.
A
ch
ar
ac
te
r
-
lev
el
m
eth
o
d
is
p
r
o
p
o
s
ed
to
g
en
er
ate
Ar
ab
ic
ad
v
e
r
s
ar
ial
ex
am
p
les
th
at
f
o
o
le
d
a
b
id
ir
ec
tio
n
al
en
co
d
er
r
ep
r
es
en
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
er
s
(
B
E
R
T
)
[
2
7
]
m
o
d
el
t
r
ain
ed
o
n
th
e
b
o
o
k
r
ev
iews
in
Ar
ab
ic
d
ataset
(
B
R
AD
)
[
2
8
]
.
T
h
e
p
r
o
p
o
s
ed
attac
k
d
ec
r
ea
s
e
d
th
e
class
if
icatio
n
ac
cu
r
ac
y
o
f
th
e
SA
class
if
ier
b
y
an
av
e
r
ag
e
o
f
3
.
4
4
%.
Ou
r
m
eth
o
d
u
s
es,
o
n
th
e
o
n
e
h
an
d
,
th
e
co
m
b
in
ed
s
co
r
e
f
u
n
ctio
n
(
C
SF
)
[
1
4
]
to
id
en
tify
th
e
m
o
s
t
in
f
lu
en
tial
to
k
en
s
in
th
e
in
p
u
t
tex
t.
T
h
e
s
ec
o
n
d
s
tep
is
to
m
o
d
if
y
th
ese
wo
r
d
s
wh
ile
p
r
eser
v
in
g
th
eir
s
im
ilar
ity
to
th
e
o
r
ig
in
al
o
n
es
co
n
s
id
er
in
g
th
e
p
ar
ticu
lar
ity
o
f
t
h
e
Ar
ab
ic
lan
g
u
a
g
e.
O
u
r
ad
v
er
s
ar
ial
ex
am
p
les
ca
n
b
e
b
u
ilt
in
B
B
s
ettin
g
s
with
o
u
t
k
n
o
wled
g
e
o
f
th
e
in
n
e
r
c
o
n
f
ig
u
r
atio
n
o
f
th
e
m
o
d
el.
Ad
d
itio
n
ally
,
it
is
s
im
p
le,
h
ar
d
ly
n
o
ticea
b
le,
an
d
r
e
q
u
ir
es
m
in
o
r
tr
a
n
s
f
o
r
m
atio
n
s
.
Fu
r
th
e
r
m
o
r
e,
to
th
e
b
est
o
f
o
u
r
k
n
o
wled
g
e
,
o
u
r
s
tu
d
y
is
t
h
e
f
ir
s
t
to
ev
alu
ate
th
e
a
d
v
er
s
ar
ial
r
o
b
u
s
tn
es
s
o
f
a
PLM
m
o
d
el
f
o
r
th
e
SA
task
,
u
s
in
g
a
ch
ar
ac
ter
-
lev
el
attac
k
,
co
m
p
ar
ed
to
p
r
io
r
r
esear
c
h
th
a
t e
x
am
in
ed
a
d
v
er
s
ar
ial
ex
am
p
l
es
in
Ar
ab
ic.
T
h
e
r
em
ai
n
d
e
r
o
f
t
h
e
p
ap
er
i
s
o
r
g
an
ize
d
as
f
o
ll
o
ws
:
s
e
cti
o
n
2
p
r
o
v
id
es
a
r
e
v
i
ew
o
f
r
e
la
ted
wo
r
k
s
.
Secti
o
n
3
p
r
es
en
ts
o
u
r
m
et
h
o
d
t
o
g
e
n
e
r
at
e
a
d
v
e
r
s
a
r
i
al
e
x
a
m
p
les
an
d
g
i
v
es
i
n
f
o
r
m
a
ti
o
n
ab
o
u
t
o
u
r
e
x
p
e
r
i
m
e
n
t
.
Secti
o
n
4
d
is
c
u
s
s
es
o
u
r
r
es
u
lts
a
n
d
p
r
ese
n
ts
s
o
m
e
g
e
n
er
ate
d
a
d
v
er
s
ar
ial
e
x
a
m
p
les
.
I
t
als
o
c
o
n
tai
n
s
co
m
p
le
m
e
n
t
ar
y
a
n
a
ly
s
es,
n
a
m
ely
co
m
p
o
n
en
t
-
w
is
e
a
n
a
ly
s
is
,
ab
l
ati
o
n
s
t
u
d
y
,
co
m
p
a
r
is
o
n
w
i
th
b
ase
li
n
e
m
et
h
o
d
,
co
m
p
a
r
is
o
n
wi
th
e
x
is
ti
n
g
A
r
a
b
ic
att
ac
k
,
a
n
d
tr
a
n
s
f
er
a
b
ili
ty
a
n
aly
s
is
.
Fi
n
a
ll
y
,
s
e
cti
o
n
5
p
r
o
v
i
d
es
c
o
n
cl
u
s
i
o
n
.
2.
RE
L
AT
E
D
WO
RK
T
h
is
s
ec
tio
n
p
r
esen
t stu
d
ies
th
at
in
v
esti
g
ate
ad
v
er
s
ar
ial
attac
k
s
in
th
e
Ar
ab
ic
L
an
g
u
ag
e.
T
h
ese
wo
r
k
s
co
n
d
u
ct
attac
k
s
em
p
l
o
y
in
g
d
i
f
f
er
en
t
a
p
p
r
o
ac
h
es
a
n
d
e
x
am
i
n
in
g
v
ar
io
u
s
co
n
tex
ts
.
T
h
en
,
T
ab
le
1
s
y
n
th
esizes
th
ese
s
tu
d
ies b
ased
o
n
th
e
y
ea
r
o
f
p
u
b
licatio
n
,
th
e
NL
P task
,
th
e
f
o
o
le
d
m
o
d
el,
an
d
th
e
atta
ck
g
r
an
u
lar
ity
.
Als
h
em
ali
an
d
Kalita
[
2
9
]
v
i
o
lated
th
e
n
o
u
n
-
ad
jectiv
e
a
g
r
ee
m
en
t
to
g
en
e
r
ate
ad
v
e
r
s
ar
ial
ex
am
p
les
.
A
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
iLST
M)
m
o
d
el
[
3
0
]
a
n
d
a
c
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN)
m
o
d
el
[
3
1
]
is
s
u
cc
ess
f
u
lly
f
o
o
led
.
As a
r
esu
lt,
th
eir
p
er
f
o
r
m
a
n
ce
s
o
n
th
e
SA ta
s
k
d
eter
io
r
at
e.
Alb
ilali
et
a
l.
[
3
2
]
ch
an
g
e
r
an
d
o
m
ly
th
e
wo
r
d
s
o
r
d
e
r
s
in
th
e
in
p
u
t.
T
h
e
y
g
en
e
r
ate
ad
v
er
s
ar
ial
ex
am
p
les
i
n
th
is
m
an
n
er
to
p
r
o
b
e
Ar
ab
ic
b
id
ir
ec
tio
n
al
en
co
d
er
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
er
s
(
Ar
aBER
T
)
[
3
3
]
,
an
Ar
a
b
ic
PLM
b
ased
o
n
B
E
R
T
.
C
o
n
s
eq
u
e
n
tly
,
t
h
e
p
er
f
o
r
m
a
n
ce
s
o
f
th
e
m
o
d
el
d
ec
r
ea
s
ed
in
th
e
m
ac
h
in
e
r
ea
d
i
n
g
co
m
p
r
e
h
en
s
io
n
(
MRC
)
task
.
Acc
o
r
d
in
g
to
Als
h
em
ali
an
d
Kalita
[
3
4
]
,
b
ased
o
n
v
is
u
al
s
im
ilar
ity
b
etwe
en
s
o
m
e
Ar
ab
ic
letter
s
,
an
alg
o
r
ith
m
to
g
e
n
er
ate
ch
ar
ac
ter
-
lev
el
ad
v
er
s
ar
ial
ex
am
p
le
s
was
s
u
g
g
ested
.
I
t
h
as
s
u
cc
ess
f
u
lly
f
o
o
led
a
B
iLST
M
m
o
d
el
an
d
a
C
NN
m
o
d
el
f
o
r
th
e
SA
task
.
B
E
R
T
an
d
XL
Net
[
3
5
]
m
o
d
els
wer
e
a
ls
o
in
v
esti
g
ated
f
o
r
th
e
task
o
f
n
ews c
ateg
o
r
izatio
n
(
NC
)
.
A
l
s
h
al
a
n
a
n
d
R
e
k
a
b
d
a
r
[
3
6
]
s
u
g
g
e
s
t
s
t
w
o
w
o
r
d
-
l
e
v
e
l
ap
p
r
o
a
c
h
e
s
t
o
t
r
i
c
k
t
h
e
A
r
aB
E
R
T
a
n
d
c
o
m
p
r
e
h
e
n
s
i
v
e
,
A
r
a
b
i
c
,
m
u
lt
i
-
d
i
a
l
ec
t
,
e
x
t
e
n
s
i
v
e
,
l
e
a
r
n
i
n
g
b
i
d
i
r
e
c
t
i
o
n
a
l
e
n
c
o
d
e
r
r
e
p
r
e
s
e
n
t
a
ti
o
n
s
f
r
o
m
t
r
a
n
s
f
o
r
m
e
r
s
(
C
A
M
e
L
B
E
R
T
)
m
o
d
e
l
s
[
3
7
]
.
T
h
e
y
s
u
b
s
t
i
t
u
t
e
th
e
w
o
r
d
s
o
n
c
e
a
t
r
a
n
d
o
m
a
n
d
a
g
a
i
n
b
a
s
e
d
o
n
t
h
e
i
r
w
e
i
g
h
t
s
.
T
h
e
r
e
f
o
r
e
,
t
h
e
a
c
c
u
r
ac
y
o
f
t
r
a
n
s
f
o
r
m
e
r
-
b
a
s
e
d
t
e
x
t
cl
a
s
s
i
f
i
e
r
s
h
as
b
e
e
n
r
e
d
u
c
e
d
.
A
ls
h
a
h
r
a
n
i
e
t
a
l
.
[
3
8
]
p
r
o
p
o
s
e
s
y
n
o
n
y
m
-
b
ased
wo
r
d
-
lev
el
attac
k
.
T
h
ey
u
s
e
B
E
R
T
m
o
d
el
t
o
g
e
n
er
ate
a
d
v
er
s
ar
ia
l
ex
am
p
les
.
T
h
ese
ad
v
er
s
ar
ial
ex
am
p
les
ar
e
em
p
l
o
y
ed
a
g
ain
s
t th
e
C
NN,
B
iLST
M
,
an
d
B
E
R
T
m
o
d
els.
Nak
h
leh
et
a
l.
[
3
9
]
ex
am
in
e
d
th
e
ad
v
e
r
s
ar
ial
r
o
b
u
s
tn
ess
o
f
t
wo
Ar
aBER
T
m
o
d
els
u
s
in
g
a
d
v
er
s
ar
ial
ex
am
p
les
b
ased
o
n
s
p
ellin
g
er
r
o
r
s
.
T
h
e
r
ep
lace
m
e
n
t
was
i
n
s
p
ir
ed
b
y
th
at
s
u
g
g
ested
b
y
th
e
s
tu
d
y
[
3
4
]
.
T
h
e
wo
r
d
to
b
e
tr
an
s
f
o
r
m
ed
was
ch
o
s
en
b
ased
o
n
th
e
B
E
R
T
s
en
tim
en
t
s
co
r
e,
wh
ich
ca
lc
u
lates
th
e
co
n
tr
i
b
u
tio
n
o
f
ea
ch
wo
r
d
to
th
e
p
r
ed
ictio
n
.
An
in
v
esti
g
atio
n
o
f
th
e
ad
v
er
s
ar
ial
r
o
b
u
s
tn
ess
o
f
a
n
Ar
ab
ic
o
f
f
en
s
iv
e
lan
g
u
ag
e
class
if
ier
wa
s
p
er
f
o
r
m
ed
in
[
4
0
]
.
An
ex
p
lain
a
b
le
ar
tific
ial
in
tellig
en
ce
(
XAI
)
ap
p
r
o
ac
h
was
em
p
lo
y
ed
to
g
en
er
ate
a
d
v
er
s
ar
ial
ex
am
p
les
.
T
h
e
Ar
aBER
T
an
d
Qata
r
C
o
m
p
u
tin
g
R
esear
ch
I
n
s
titu
te
Ar
ab
ic
an
d
D
ialec
tal
(
QARi
B
)
m
o
d
els we
r
e
f
o
o
led
b
y
th
ese
ad
v
e
r
s
ar
ial
ex
am
p
les.
T
ab
le
1
.
R
elate
d
wo
r
k
S
t
u
d
y
Y
e
a
r
N
LP t
a
s
k
M
o
d
e
l
A
t
t
a
c
k
g
r
a
n
u
l
a
r
i
t
y
[
2
9
]
2
0
1
9
SA
C
N
N
,
B
i
LST
M
C
h
a
r
a
c
t
e
r
-
l
e
v
e
l
[
3
2
]
2
0
2
1
M
R
C
A
r
a
B
ER
T
W
o
r
d
-
l
e
v
e
l
[
3
4
]
2
0
2
1
SA
C
N
N
,
B
i
LST
M
C
h
a
r
a
c
t
e
r
-
l
e
v
e
l
[
3
5
]
2
0
1
9
NC
B
ER
T,
X
LN
e
t
C
h
a
r
a
c
t
e
r
-
l
e
v
e
l
[
3
6
]
2
0
2
3
Te
x
t
c
l
a
ssi
f
i
c
a
t
i
o
n
A
r
a
B
ER
T,
C
a
m
e
l
B
E
R
T
W
o
r
d
-
l
e
v
e
l
[
3
8
]
2
0
2
4
SA
C
N
N
,
B
i
LST
M
,
B
ER
T
W
o
r
d
-
l
e
v
e
l
[
3
9
]
2
0
2
4
SA
A
r
a
B
ER
T
C
h
a
r
a
c
t
e
r
-
l
e
v
e
l
[
4
0
]
2
0
2
4
Te
x
t
c
l
a
ssi
f
i
c
a
t
i
o
n
A
r
a
B
ER
T,
Q
A
R
i
B
W
o
r
d
-
l
e
v
e
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
d
ve
r
s
a
r
ia
l e
xa
mp
les in
A
r
a
b
ic
la
n
g
u
a
g
e
(
S
a
fa
e
La
a
tya
o
u
i
)
943
3.
M
E
T
H
O
D
Af
ter
p
r
esen
tin
g
s
ev
er
al
p
ap
e
r
s
th
at
p
r
o
v
id
e
s
u
g
g
esti
o
n
s
f
o
r
g
e
n
er
atin
g
ad
v
er
s
ar
ial
ex
a
m
p
les
,
th
is
s
ec
tio
n
d
escr
ib
es
th
e
p
r
o
p
o
s
e
d
m
eth
o
d
.
I
t
is
a
two
-
s
tep
m
eth
o
d
to
g
en
er
ate
ad
v
er
s
ar
ial
ex
am
p
l
es
in
Ar
ab
ic
lan
g
u
ag
e:
f
ir
s
t,
a
s
co
r
e
is
ca
l
cu
lated
f
o
r
ea
ch
to
k
en
b
ased
o
n
its
ef
f
ec
t
o
n
th
e
m
o
d
el
p
r
ed
ictio
n
,
an
d
t
h
e
s
eq
u
en
ce
to
k
en
s
ar
e
r
an
k
ed
a
cc
o
r
d
in
g
to
th
is
s
co
r
e.
T
h
en
,
th
e
m
o
s
t
im
p
o
r
tan
t
t
o
k
en
s
a
r
e
tr
an
s
f
o
r
m
e
d
wh
ile
p
r
eser
v
in
g
t
h
eir
s
im
ilar
ity
to
t
h
e
in
itial o
n
es.
3
.
1
.
T
o
k
en
s
co
ring
a
nd
ra
nk
ing
Fig
u
r
e
1
s
u
m
m
ar
izes
th
e
f
ir
s
t
s
tep
o
f
th
e
s
u
g
g
ested
m
eth
o
d
.
T
h
e
im
p
o
r
tan
ce
o
f
ea
ch
to
k
en
in
th
e
s
eq
u
en
ce
is
esti
m
ated
.
T
h
e
to
k
en
s
ar
e
th
en
r
a
n
k
ed
b
ased
o
n
t
h
eir
im
p
o
r
tan
ce
.
Fig
u
r
e
1
.
I
ll
u
s
tr
atio
n
o
f
s
tep
1
T
o
ev
alu
ate
th
e
im
p
o
r
tan
ce
o
f
ea
ch
to
k
en
in
th
e
in
p
u
t
tex
t
s
eq
u
en
ce
,
th
e
C
SF
f
u
n
ctio
n
is
u
s
ed
[
1
4
]
.
E
ac
h
to
k
en
’
s
s
co
r
e
is
d
eter
m
in
ed
b
y
t
h
is
f
u
n
ctio
n
.
I
t is d
ef
in
e
d
as in
(
1
)
.
(
)
=
(
)
+
(
(
)
(
1
)
W
h
e
r
e
i
s
t
h
e
i
t
h
t
o
k
e
n
o
f
t
h
e
in
p
u
t
t
e
x
t
s
e
q
u
e
n
c
e
.
(
)
i
s
t
h
e
t
e
m
p
o
r
a
l
h
e
a
d
s
c
o
r
e
o
f
t
h
e
it
h
t
o
k
e
n
.
I
t
i
s
t
h
e
d
i
f
f
e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
m
o
d
e
l
’
s
p
r
e
d
i
c
ti
o
n
s
c
o
r
e
as
i
t
r
e
a
d
s
th
e
s
e
q
u
e
n
c
e
u
p
t
o
t
h
e
it
h
t
o
k
en
,
a
n
d
t
h
e
m
o
d
e
l
’
s
p
r
e
d
i
c
t
i
o
n
s
c
o
r
e
a
s
it
r
e
a
d
s
i
t
u
p
t
o
t
h
e
(
−
1
)
t
h
t
o
k
e
n
.
I
t
r
e
f
l
e
c
t
s
t
h
e
i
n
f
l
u
e
n
c
e
o
f
t
h
e
t
o
k
en
o
n
t
h
e
f
i
n
a
l
p
r
e
d
i
c
t
i
o
n
w
h
e
n
c
o
u
p
l
e
d
w
it
h
th
e
t
o
k
e
n
s
t
h
a
t
p
r
e
c
e
d
e
i
t
.
F
b
e
in
g
t
h
e
t
a
r
g
e
t
e
d
m
o
d
e
l
,
TH
S
i
s
ca
l
c
u
l
a
t
e
d
as
i
n
(
2
)
.
(
)
=
(
1
,
2
,
.
.
.
,
−
1
,
)
−
(
1
,
2
,
.
.
.
,
−
1
)
(
2
)
W
h
e
r
e
is
a
h
y
p
er
p
ar
am
eter
.
I
n
th
is
ca
s
e,
λ
=1
is
u
s
ed
,
g
iv
i
n
g
th
e
s
am
e
im
p
o
r
tan
ce
to
th
e
two
ter
m
s
:
T
HS
an
d
T
T
S.
(
)
is
th
e
tem
p
o
r
al
tail
s
co
r
e
o
f
th
e
i
-
th
to
k
en
.
I
t
is
th
e
d
if
f
er
en
ce
b
etwe
en
two
tr
ailin
g
p
ar
ts
o
f
th
e
s
eq
u
en
ce
,
o
n
e
co
n
tain
in
g
th
e
ith
t
o
k
en
wh
ile
t
h
e
o
th
er
d
o
es
n
o
t.
I
t
r
ef
lects
t
h
e
in
f
lu
e
n
ce
o
f
th
e
to
k
en
o
n
th
e
f
in
al
p
r
e
d
ictio
n
wh
e
n
co
u
p
led
with
to
k
en
s
th
at
f
o
llo
wed
i
t.
I
t is co
m
p
u
ted
ac
co
r
d
i
n
g
to
(
3
)
.
(
)
=
(
,
+
1
,
.
.
.
,
−
1
,
)
−
(
+
1
,
.
.
.
,
−
1
,
)
(
3
)
As
s
ee
n
in
its
d
ef
in
itio
n
,
th
is
s
co
r
e
f
u
n
ctio
n
is
ad
ap
te
d
to
B
B
s
ettin
g
s
.
T
h
is
i
s
b
ec
au
s
e
it
r
elies
o
n
ly
o
n
th
e
p
r
e
d
ictio
n
o
f
th
e
in
iti
al
m
o
d
el.
On
ce
th
e
to
k
en
s
ar
e
s
co
r
ed
an
d
r
an
k
ed
,
th
e
n
ex
t
s
tep
,
n
am
ely
th
e
tr
an
s
f
o
r
m
atio
n
,
is
p
er
f
o
r
m
ed
.
3
.
2
.
T
o
k
en
t
ra
ns
f
o
r
m
er
Ou
r
tr
an
s
f
o
r
m
atio
n
is
ch
ar
ac
te
r
-
lev
el.
I
t a
f
f
ec
ts
ju
s
t o
n
e
ch
ar
ac
ter
o
f
th
e
m
o
s
t im
p
o
r
tan
t w
o
r
d
s
o
f
th
e
in
p
u
t
tex
t
s
eq
u
en
ce
.
Fig
u
r
e
2
s
u
m
m
ar
izes
th
e
s
ec
o
n
d
s
tep
o
f
th
e
s
u
g
g
ested
m
et
h
o
d
,
wh
ic
h
co
n
s
is
ts
o
f
t
h
r
ee
v
ar
ian
ts
:
Ar
-
Attack
1
as
in
Fig
u
r
e
2
(
a)
,
Ar
-
Attack
2
as
in
Fig
u
r
e
2
(
b
)
,
an
d
Ar
-
Attack
5
as
in
Fig
u
r
e
2
(
c)
.
T
h
e
tr
an
s
f
o
r
m
atio
n
alg
o
r
ith
m
ap
p
li
es o
n
e
o
f
th
e
two
m
o
d
if
icatio
n
s
lis
ted
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
941
-
9
5
2
944
(
a)
(
b
)
(
c)
Fig
u
r
e
2
.
I
ll
u
s
tr
atio
n
o
f
s
tep
2
o
f
(
a)
Ar
-
Attack
1
,
(
b
)
Ar
-
Attack
2
,
an
d
(
c)
Ar
-
Attack
5
3
.
2
.
1
.
M
o
difica
t
io
n 1
:
re
pla
c
em
ent
o
f
o
ne
cha
ra
ct
er
Sev
er
al
Ar
ab
ic
letter
s
s
o
u
n
d
e
x
tr
em
ely
s
im
ilar
,
b
u
t
s
o
m
e
ar
e
mu
r
a
q
a
q
a
an
d
o
th
e
r
s
ar
e
m
u
fa
kh
a
ma
.
T
ab
le
2
p
r
esen
ts
a
co
m
p
ilati
o
n
o
f
letter
s
ca
teg
o
r
ized
as
mu
r
a
q
a
q
a
an
d
t
h
eir
co
r
r
esp
o
n
d
in
g
c
o
u
n
ter
p
ar
ts
class
if
ied
as
mu
fa
kh
a
ma
,
ac
c
o
m
p
an
ied
b
y
th
eir
r
esp
ec
tiv
e
in
ter
n
atio
n
al
p
h
o
n
etic
alp
h
a
b
et
(
I
PA)
n
o
tatio
n
s
en
clo
s
ed
in
s
id
e
p
ar
en
th
eses
.
Fo
r
th
is
m
o
d
if
icatio
n
,
th
e
mu
r
a
q
a
q
a
c
h
ar
ac
ter
is
r
ep
lace
d
with
its
co
r
r
esp
o
n
d
in
g
mu
fa
kh
a
ma
o
n
e
an
d
v
ice
v
er
s
a.
T
ab
le
2
.
T
h
e
mu
r
a
q
a
q
a
letter
s
an
d
th
e
c
o
r
r
esp
o
n
d
in
g
mu
fa
k
h
a
ma
o
n
es
Mu
r
a
q
a
q
a
c
h
a
r
a
c
t
e
r
Mu
f
a
k
h
a
m
a
c
h
a
r
a
c
t
e
r
س
(
s)
ص
(s
ʕ
)
د
(
d
)
ض
(d
ʕ
)
ت
(
t
)
ط
(t
ʕ
)
3
.
2
.
2
.
M
o
difica
t
io
n 2
:
a
dd
it
io
n o
f
o
ne
cha
ra
c
t
er
A
lif
A
lma
d
is
a
s
p
ec
ial
ch
ar
ac
ter
in
Ar
ab
ic.
So
m
etim
es
it
is
wr
itten
with
o
u
t
b
ein
g
r
ea
d
,
an
d
o
th
e
r
tim
es
it
is
r
ea
d
with
o
u
t
b
ein
g
wr
itten
.
T
ab
le
3
s
h
o
ws
s
o
m
e
wo
r
d
s
th
at
co
n
tain
A
lif
A
lma
d
an
d
h
o
w
r
ea
d
th
em
.
T
h
e
f
ir
s
t
co
lu
m
n
p
r
esen
ts
h
o
w
to
wr
ite
th
em
a
n
d
th
e
s
ec
o
n
d
o
n
e
p
r
esen
ts
h
o
w
to
r
ea
d
th
em
.
Fu
r
th
er
m
o
r
e,
its
s
o
u
n
d
is
v
er
y
s
im
ilar
to
th
at
o
f
th
e
d
iacr
itical
m
ar
k
A
lfa
th
a
.
T
ab
le
3
.
W
o
r
d
s
co
n
tain
in
g
t
h
e
A
lif A
lma
d
ch
ar
ac
ter
W
o
r
d
s
H
o
w
t
o
r
e
a
d
t
h
e
m
ك
لذ
ك
(
Li
k
e
t
h
a
t
)
ك
لا
ذ
ك
اذ
ه
(
T
h
i
s)
اذ
ا
ه
هذ
ه
(
T
h
i
s
–
f
o
r
t
h
e
f
e
m
i
n
i
n
e
)
هذ
ا
ه
اول
م
ع
(
Th
e
y
w
o
r
k
e
d
)
ولم
ع
اوس
ر
د
(
T
h
e
y
s
t
u
d
i
e
d
)
وس
ر
د
T
ab
le
4
p
r
esen
ts
,
in
th
e
f
ir
s
t
lin
e,
th
e
A
lif
A
lma
d
‘s
I
PA
n
o
tatio
n
.
I
t
p
r
esen
ts
,
in
th
e
s
ec
o
n
d
lin
e,
th
e
I
PA n
o
tatio
n
o
f
th
e
d
iacr
itical
m
ar
k
A
lfa
th
a
.
Fo
r
th
is
s
ec
o
n
d
m
o
d
if
icatio
n
,
A
lif A
lma
d
is
p
r
o
p
o
s
ed
to
b
e
a
d
d
ed
to
th
e
en
d
o
f
th
e
m
o
d
i
f
ied
wo
r
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
d
ve
r
s
a
r
ia
l e
xa
mp
les in
A
r
a
b
ic
la
n
g
u
a
g
e
(
S
a
fa
e
La
a
tya
o
u
i
)
945
T
ab
le
4
.
A
lif A
lma
d
a
n
d
A
lfa
t
h
a
’
s
I
PA n
o
tatio
n
C
h
a
r
a
c
t
e
r
I
P
A
n
o
t
a
t
i
o
n
Al
i
f
A
l
m
a
d
a
Al
f
a
t
h
a
a:
3
.
2
.
3
.
T
ra
ns
f
o
rm
a
t
i
o
n
a
lg
o
ri
t
hm
Fo
r
th
e
to
k
en
th
at
will
b
e
t
r
a
n
s
f
o
r
m
ed
,
th
e
al
g
o
r
ith
m
p
r
ese
n
ted
in
Fig
u
r
e
3
is
ap
p
lied
.
I
f
th
e
to
k
en
th
at
will
b
e
tr
an
s
f
o
r
m
ed
d
o
es
n
o
t
co
n
tain
an
y
ch
a
r
ac
ter
f
r
o
m
th
e
T
ar
g
etC
h
ar
lis
t,
th
e
A
lif
A
lma
d
ch
ar
ac
ter
is
ad
d
ed
at
th
e
e
n
d
o
f
t
h
e
wo
r
d
.
I
f
it
co
n
tain
s
ju
s
t
o
n
e
ch
ar
ac
te
r
am
o
n
g
th
e
ch
ar
ac
te
r
s
o
n
th
e
lis
t,
th
is
ch
ar
ac
ter
will b
e
r
ep
lace
d
with
th
e
co
r
r
e
s
p
o
n
d
in
g
o
n
e
as d
ef
in
e
d
in
T
a
b
le
2
.
I
f
it c
o
n
tain
s
m
an
y
ch
a
r
ac
ter
s
f
r
o
m
th
e
lis
t,
a
s
in
g
le
ch
ar
ac
ter
is
r
an
d
o
m
ly
s
elec
ted
an
d
r
ep
lace
d
in
th
e
s
a
m
e
way
.
Fig
u
r
e
3
.
T
r
an
s
f
o
r
m
ati
o
n
alg
o
r
ith
m
3
.
3
.
P
r
o
po
s
ed
a
dv
er
s
a
ria
l a
t
t
a
ck
s
T
h
is
wo
r
k
p
r
o
p
o
s
es
th
r
ee
ad
v
er
s
ar
ial
attac
k
s
:
Ar
-
Attack
1
,
Ar
-
Attack
2
,
an
d
Ar
-
Attack
5
.
Fo
r
th
e
Ar
-
Attack
1
,
th
e
ab
o
v
e
tr
an
s
f
o
r
m
atio
n
alg
o
r
ith
m
is
ap
p
lied
t
o
o
n
e
to
k
e
n
:
th
e
m
o
s
t
im
p
o
r
tan
t
to
k
en
b
ased
o
n
th
e
C
SF
.
Fo
r
th
e
Ar
-
Attack
2
,
th
e
ab
o
v
e
tr
an
s
f
o
r
m
atio
n
al
g
o
r
ith
m
is
a
p
p
lied
t
o
two
t
o
k
en
s
:
th
e
f
ir
s
t
a
n
d
s
ec
o
n
d
two
m
o
s
t
im
p
o
r
tan
t
o
n
es.
Fo
r
th
e
Ar
-
Attack
5
,
th
e
ab
o
v
e
tr
a
n
s
f
o
r
m
atio
n
alg
o
r
ith
m
is
ap
p
lied
to
f
i
v
e
to
k
en
s
:
th
e
f
iv
e
m
o
s
t
im
p
o
r
tan
t
o
n
es.
Fro
m
o
n
e
in
p
u
t
te
x
t,
th
r
ee
ad
v
e
r
s
ar
ial
ex
am
p
le
s
ar
e
cr
ea
ted
.
An
ex
am
p
le
is
s
h
o
wn
in
T
a
b
le
5
.
T
ab
le
5
.
Dif
f
e
r
en
t
ad
v
er
s
ar
ial
ex
am
p
les g
en
er
ate
d
b
ased
o
n
an
in
p
u
t te
x
t seq
u
en
ce
O
r
i
g
i
n
a
l
i
n
p
u
t
t
e
x
t
ي
ف
ةلا
م
ع
لا
ةي
ض
ق
ل
ض
ر
ع
ت
هن
أ
ي
ن
ب
ج
ع
أ
ا
م
ر
ث
ك
أ
ك
ا
ن
ه ع
ا
ض
و
ا
ي
ف
ك
ش
أ
ت
ن
ك
ا
م
لا
ط
.
ء
ا
ح
ط
ب
لا
(
W
h
a
t
I
f
o
u
n
d
i
n
t
e
r
e
st
i
n
g
w
a
s
t
h
a
t
h
e
w
a
s
e
x
p
o
se
d
t
o
t
h
e
l
a
b
o
r
i
ss
u
e
i
n
A
l
b
a
t
h
a
.
A
s l
o
n
g
a
s I
d
o
u
b
t
e
d
t
h
e
s
i
t
u
a
t
i
o
n
t
h
e
r
e
.
)
Ar
-
A
t
t
a
c
k
1
ب
لا
ي
ف
ةلا
م
ع
لا
ةي
ض
ق
ل
ض
ر
ع
ت
هن
أ
ي
ن
ب
ج
ع
أ
ا
م
ر
ث
ك
أ
ت
ي
ف
ك
ش
أ
ت
ن
ك
ا
م
لا
ط
.
ء
ا
ح
ك
ا
ن
ه ع
ا
ض
و
ا
Ar
-
A
t
t
a
c
k
2
ا
م
ر
ث
ك
أ
ا
ض
و
ا
ي
ف
ك
ش
أ
ت
ن
ك
ا
م
لا
ط
.
ء
ا
ح
ت
ب
لا
ي
ف
ةلا
م
ع
لا
ةي
د
ق
ل
ض
ر
ع
ت
هن
أ
ي
ن
ب
ج
ع
أ
ك
ا
ن
ه ع
Ar
-
A
t
t
a
c
k
5
ر
ث
ك
أ
د
و
ا
ي
ف
ك
ش
أ
طن
ك
ا
م
لا
ط
.
ء
ا
ح
ت
ب
لا
ي
ف
ةلا
م
ع
ل
ا
ةي
د
ق
ل
ض
ر
ع
ت
هن
أ
ي
ن
ب
ج
ع
أ
اا
م
ك
ا
ن
ه ع
ا
3
.
4
.
Sim
ila
ri
t
y
is
s
ue
T
h
e
d
if
f
er
en
ce
b
etwe
e
n
th
e
o
r
ig
in
al
in
p
u
t
an
d
th
e
ad
v
er
s
ar
ial
in
p
u
t
s
h
o
u
ld
n
o
t
b
e
as
p
er
ce
p
tib
le
as
p
o
s
s
ib
le
wh
ile
cr
ea
tin
g
an
a
d
v
er
s
ar
ial
ex
a
m
p
le
.
Fo
r
th
is
r
ea
s
o
n
,
th
e
ap
p
lied
m
o
d
if
icat
io
n
s
as
d
etailed
in
s
u
b
-
s
ec
tio
n
s
3
.
2
.
1
an
d
3
.
2
.
2
ar
e
ca
r
ef
u
lly
ch
o
s
en
:
th
e
o
r
ig
in
al
ch
a
r
ac
ter
a
n
d
th
e
m
o
d
if
ied
o
n
e
s
o
u
n
d
r
em
ar
k
ab
l
y
s
im
ilar
,
a
n
d
t
h
ese
m
o
d
if
icatio
n
s
p
r
esen
t
co
m
m
o
n
m
is
tak
es
in
Ar
ab
ic
w
r
itin
g
[
4
1
]
.
T
h
is
m
ak
es
it
p
o
s
s
ib
le
f
o
r
ad
v
er
s
ar
ial
e
x
am
p
les to
g
o
u
n
n
o
ticed
o
r
f
o
r
th
e
r
ea
d
er
to
th
in
k
it is
s
im
p
ly
a
s
p
ellin
g
er
r
o
r
.
T
o
m
ea
s
u
r
e
h
o
w
s
im
ilar
th
e
g
en
er
ated
ad
v
e
r
s
ar
ial
ex
am
p
les
ar
e
to
th
e
o
r
ig
in
al
in
p
u
ts
,
th
e
n
u
m
b
er
o
f
ch
an
g
es
m
etr
ic
is
u
s
ed
.
I
t
is
an
ed
it
-
b
ased
m
ea
s
u
r
em
en
t,
a
way
o
f
q
u
an
tif
y
in
g
c
h
an
g
es
f
r
o
m
o
n
e
s
tr
in
g
t
o
an
o
th
er
[
8
]
.
I
t e
q
u
als 1
,
2
,
a
n
d
5
f
o
r
A
r
-
attac
k
1
,
Ar
-
attac
k
2
,
a
n
d
Ar
-
attac
k
5
r
esp
ec
tiv
ely
.
3
.
5
.
E
x
perim
ent
Ou
r
ad
v
e
r
s
ar
ial
attac
k
s
wer
e
im
p
lem
en
ted
u
s
in
g
Py
th
o
n
o
n
Go
o
g
le
C
o
lab
o
r
ato
r
y
.
T
h
e
ex
p
er
im
en
t
was
ap
p
lied
t
o
a
b
in
a
r
y
class
if
ier
b
ased
o
n
B
E
R
T
an
d
tr
ai
n
ed
o
n
th
e
B
R
AD
d
ataset
f
o
r
th
e
SA
task
.
T
h
e
p
er
f
o
r
m
an
ce
was e
v
alu
ated
u
s
in
g
ac
cu
r
ac
y
m
etr
ic.
T
h
e
test
in
g
s
tr
ateg
y
is
also
d
escr
ib
ed
in
th
is
s
u
b
-
s
ec
tio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
941
-
9
5
2
946
3
.
5
.
1
.
Wo
rk
s
pa
ce
a
nd
lib
ra
ries
C
o
lab
o
r
ato
r
y
,
o
r
C
o
lab
f
o
r
s
h
o
r
t,
is
a
p
r
o
d
u
ct
f
r
o
m
Go
o
g
le
R
esear
ch
.
C
o
lab
allo
ws
y
o
u
to
wr
ite
an
d
ex
ec
u
te
Py
th
o
n
co
d
e
th
r
o
u
g
h
th
e
b
r
o
wser
.
T
o
b
en
e
f
it
f
r
o
m
b
etter
r
eso
u
r
ce
s
th
an
th
o
s
e
g
iv
en
b
y
th
e
f
r
ee
v
er
s
io
n
o
f
C
o
lab
,
th
e
p
aid
v
e
r
s
io
n
C
o
lab
Pro
was
u
s
ed
.
A
v
ar
iety
o
f
Py
th
o
n
lib
r
ar
ies
wer
e
u
s
ed
:
Pan
d
as,
Nu
m
Py
,
Scik
it
-
lear
n
,
T
e
n
s
o
r
f
l
o
w,
Ker
as,
an
d
tr
a
n
s
f
o
r
m
e
r
s
.
3
.
5
.
2
.
Da
t
a
s
et
T
h
e
b
alan
ce
d
v
e
r
s
io
n
o
f
B
R
A
D
is
u
s
ed
.
A
lar
g
e
tex
tu
al
d
at
aset
f
o
r
th
e
SA
task
th
at
co
n
t
ain
s
m
o
r
e
th
an
1
5
6
,
0
0
0
r
ev
iews
co
llected
f
r
o
m
th
e
G
o
o
d
R
ea
d
s
.
co
m
web
s
ite
in
2
0
1
6
.
T
h
e
r
e
v
iews
wer
e
wr
itten
m
ain
ly
in
s
tan
d
ar
d
Ar
ab
ic
b
u
t
also
u
s
in
g
d
ialec
tal
Ar
ab
ic.
Fiv
e
co
lu
m
n
s
co
m
p
o
s
e
th
is
d
atas
et:
u
s
er
_
id
,
b
o
o
k
_
id
,
r
ev
iew,
r
ev
iew_
id
,
an
d
r
atin
g
.
3
.
5
.
3
.
Da
t
a
prepro
ce
s
s
ing
W
e
ar
e
in
ter
ested
o
n
ly
in
r
ev
i
ews
an
d
th
ei
r
co
r
r
esp
o
n
d
in
g
r
atin
g
s
.
R
ev
iews
in
th
e
b
alan
c
ed
v
e
r
s
io
n
ar
e
m
ap
p
ed
to
f
o
u
r
r
atin
g
v
alu
es:
1
o
r
2
f
o
r
th
e
n
eg
ati
v
e
s
en
tim
en
t
an
d
4
o
r
5
f
o
r
th
e
p
o
s
itiv
e
o
n
e.
T
o
ex
p
er
im
en
t
with
a
b
in
ar
y
c
lass
if
icatio
n
,
two
class
es
is
cr
ea
ted
:
a
n
e
g
ativ
e
class
th
at
c
o
r
r
esp
o
n
d
s
t
o
r
atin
g
s
1
an
d
2
,
an
d
a
p
o
s
itiv
e
class
t
h
at
co
r
r
esp
o
n
d
s
to
r
atin
g
s
4
a
n
d
5
.
T
h
e
d
ata
was
s
p
lit
in
to
s
u
b
s
ets
as
f
o
llo
ws:
6
0
%
f
o
r
th
e
tr
ain
in
g
s
et,
2
0
%
f
o
r
th
e
v
alid
atio
n
s
et,
an
d
2
0
%
f
o
r
th
e
test
s
et.
Du
e
to
co
m
p
u
tatio
n
al
co
n
s
tr
ain
ts
,
tr
u
n
ca
tio
n
was
ap
p
lied
,
r
esu
ltin
g
in
an
av
e
r
ag
e
in
p
u
t
le
n
g
th
o
f
ap
p
r
o
x
i
m
ately
3
3
wo
r
d
s
.
T
h
is
is
co
m
p
ar
ab
le
t
o
th
at
o
f
s
ev
er
al
wid
ely
u
s
ed
d
atasets
f
o
r
im
p
le
m
en
tin
g
attac
k
s
ag
ain
s
t b
in
a
r
y
SA
m
o
d
els.
3
.
5
.
4
.
M
o
del
Ou
r
s
u
g
g
ested
attac
k
s
wer
e
ap
p
lied
to
a
class
if
icatio
n
m
o
d
el
f
in
etu
n
in
g
B
E
R
T
,
tr
ain
ed
o
n
th
e
B
R
A
D
d
ata
s
et
as d
escr
ib
ed
in
s
u
b
-
s
ec
tio
n
s
3
.
5
.
2
an
d
3
.
5
.
3
f
o
r
1
0
ep
o
ch
s
.
T
h
e
p
r
et
r
ain
ed
m
o
d
el
an
d
to
k
en
ize
r
ar
e
b
er
t
-
b
ase
-
u
n
ca
s
ed
f
o
r
T
e
n
s
o
r
Flo
w.
T
h
e
o
p
tim
izer
is
Ad
am
1
2
.
T
h
e
lear
n
in
g
r
ate
eq
u
als
to
5
e
-
0
5
.
T
h
e
b
atch
s
iz
e
is
3
2
.
3
.
5
.
5
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n
T
o
ass
ess
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
ad
v
er
s
ar
ial
attac
k
s
,
th
e
class
if
icatio
n
ac
cu
r
ac
y
m
etr
ic
was
in
itially
u
s
ed
.
I
t
is
co
m
p
u
ted
b
ef
o
r
e
a
n
d
af
ter
attac
k
i
n
g
t
h
e
m
o
d
el
to
q
u
an
tify
th
e
o
v
er
all
d
eg
r
ad
ati
o
n
in
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
class
if
icatio
n
ac
cu
r
ac
y
is
d
ef
in
ed
as in
(
4
)
.
=
+
+
+
+
(
4
)
T
r
u
e
p
o
s
itiv
e
s
(
T
P)
an
d
tr
u
e
n
eg
ativ
es
(
T
N)
ar
e
co
r
r
ec
tly
p
r
ed
icted
p
o
s
itiv
e
a
n
d
n
eg
ativ
e
in
p
u
ts
,
r
esp
ec
tiv
ely
,
wh
ile
f
alse
p
o
s
itiv
es
(
FP
)
an
d
f
alse
n
e
g
ativ
es
(
FN)
ar
e
in
co
r
r
ec
t
p
r
ed
ictio
n
s
.
I
n
o
t
h
e
r
wo
r
d
s
,
ac
c
u
r
ac
y
m
ea
s
u
r
es th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
p
r
ed
icted
in
p
u
ts
am
o
n
g
all
in
p
u
ts
.
P
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
h
av
e
also
b
ee
n
in
clu
d
ed
,
b
e
f
o
r
e
an
d
af
ter
attac
k
,
to
p
r
o
v
i
d
e
a
m
o
r
e
d
etailed
an
aly
s
is
o
f
th
e
m
is
class
if
icatio
n
s
in
tr
o
d
u
ce
d
b
y
th
e
ad
v
er
s
ar
ial
attac
k
.
T
h
ese
m
etr
ics
ar
e
d
ef
in
e
d
as
in
(
5
)
t
o
(
7
)
.
=
+
(
5
)
=
+
(
6
)
1
−
=
2
×
×
+
(
7
)
3
.
5
.
6
.
T
esting
s
t
ra
t
eg
y
B
y
s
elec
tin
g
v
ar
io
u
s
in
p
u
ts
a
n
d
f
o
llo
win
g
th
e
m
g
r
ad
u
ally
,
th
e
p
r
o
g
r
a
m
was
test
ed
f
o
r
th
e
th
r
ee
s
tep
s
.
Fo
r
th
e
s
co
r
in
g
s
tep
,
b
y
m
ak
in
g
s
u
r
e
th
e
v
ar
i
o
u
s
s
co
r
in
g
ca
lcu
latio
n
s
ar
e
p
er
f
o
r
m
e
d
co
r
r
ec
tly
.
Fo
r
th
e
r
an
k
in
g
s
tep
,
b
y
e
n
s
u
r
in
g
th
at
th
e
tr
an
s
f
o
r
m
atio
n
s
ar
e
ap
p
li
ed
to
th
e
m
o
s
t
im
p
o
r
tan
t
to
k
e
n
b
ased
o
n
t
h
e
C
SF
f
o
r
Ar
-
Attack
1
,
th
e
f
ir
s
t
an
d
s
ec
o
n
d
two
m
o
s
t
im
p
o
r
ta
n
t
to
k
en
s
f
o
r
Ar
-
Attack
2
,
an
d
th
e
f
iv
e
m
o
s
t
im
p
o
r
tan
t
o
n
es
f
o
r
A
r
-
Attack
5
.
Fo
r
th
e
tr
an
s
f
o
r
m
in
g
s
tep
,
b
y
en
s
u
r
i
n
g
th
at
th
e
tr
an
s
f
o
r
m
atio
n
al
g
o
r
ith
m
is
co
r
r
ec
tly
ap
p
lied
to
th
e
to
k
en
th
at
n
ee
d
s
to
b
e
tr
an
s
f
o
r
m
ed
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
th
is
s
ec
tio
n
,
s
o
m
e
g
en
e
r
ated
ad
v
er
s
ar
ial
ex
am
p
les
ar
e
f
ir
s
t
p
r
esen
ted
,
alo
n
g
with
a
d
escr
ip
tio
n
o
f
h
o
w
th
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
s
u
g
g
ested
attac
k
s
o
n
th
e
m
o
d
el
in
tr
o
d
u
ce
d
in
s
ec
tio
n
3
is
ev
a
lu
ated
.
T
h
e
r
esu
lts
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
d
ve
r
s
a
r
ia
l e
xa
mp
les in
A
r
a
b
ic
la
n
g
u
a
g
e
(
S
a
fa
e
La
a
tya
o
u
i
)
947
o
f
th
e
ex
p
e
r
im
en
t
ar
e
th
e
n
s
u
m
m
ar
ized
,
in
clu
d
in
g
co
m
p
lem
en
tar
y
a
n
aly
s
es
to
f
u
r
th
er
ex
p
lo
r
e
th
e
ch
ar
ac
ter
is
tics
an
d
im
p
ac
t o
f
th
e
attac
k
s
.
Fin
ally
,
a
d
is
cu
s
s
io
n
o
f
th
e
f
in
d
i
n
g
s
is
p
r
o
v
i
d
ed
.
4
.
1
.
Sa
m
ples
o
f
g
ener
a
t
ed
a
dv
er
s
a
ria
l e
x
a
m
ples
T
ab
le
6
p
r
esen
ts
an
e
x
am
p
le
o
f
th
e
tex
t
s
eq
u
e
n
ce
o
u
tp
u
t
p
r
o
d
u
ce
d
wh
en
a
p
p
ly
in
g
th
e
a
d
v
er
s
ar
ial
attac
k
Ar
-
Attack
1
.
T
h
e
f
ir
s
t
co
lu
m
n
s
h
o
ws
th
e
o
r
ig
i
n
al
in
p
u
t
an
d
th
e
a
d
v
er
s
ar
ial
e
x
a
m
p
le
.
T
h
e
s
ec
o
n
d
co
lu
m
n
s
h
o
ws th
eir
m
o
d
el’
s
f
i
n
al
p
r
ed
ictio
n
.
T
ab
le
7
s
h
o
ws
an
ex
am
p
le
o
f
th
e
te
x
t
s
eq
u
en
ce
o
u
tp
u
t
p
r
o
d
u
ce
d
u
n
d
er
ad
v
er
s
ar
ial
a
ttack
Ar
-
Attack
2
.
T
h
e
f
ir
s
t
co
lu
m
n
co
n
tain
s
th
e
o
r
ig
in
al
in
p
u
t
an
d
th
e
ad
v
er
s
ar
ial
ex
am
p
le.
T
h
e
s
ec
o
n
d
co
lu
m
n
s
h
o
ws th
eir
m
o
d
el’
s
f
in
al
p
r
ed
ictio
n
.
T
ab
le
8
illu
s
tr
ates
an
ex
am
p
le
o
f
th
e
tex
t
s
eq
u
en
ce
o
u
tp
u
t
p
r
o
d
u
ce
d
u
n
d
er
ad
v
e
r
s
ar
ial
attac
k
Ar
-
Attack
5
.
T
h
e
f
ir
s
t
co
lu
m
n
co
n
tain
s
th
e
o
r
ig
in
al
in
p
u
t
an
d
th
e
ad
v
er
s
ar
ial
ex
am
p
le
.
T
h
e
s
ec
o
n
d
co
lu
m
n
r
ep
o
r
ts
th
eir
m
o
d
el’
s
f
in
al
p
r
e
d
ictio
n
.
T
ab
le
6
.
An
a
d
v
er
s
ar
ial
ex
a
m
p
le
g
en
er
ated
wh
en
a
p
p
ly
in
g
Ar
-
Attack
1
Te
x
t
P
r
e
d
i
c
t
i
o
n
ي
ن
ب
ج
ع
ت
م
ل
ا
ه
ن
ك
ل
و
م
و
ج
ن
ةث
ث
ق
ح
ت
س
ت
ت
ن
ا
ك
ا
م
ب
ر
0
(
Ev
e
n
t
h
o
u
g
h
i
t
ma
y
d
e
s
e
r
v
e
t
h
r
e
e
st
a
r
s;
I
w
a
s n
o
t
A
maz
e
d
b
y
i
t
.
)
م
ل
ا
ه
ن
ك
ل
و
م
و
ج
ن
ةث
ث
ق
ح
ت
س
ت
ت
ن
ا
ك
ا
م
ب
ر
ج
ع
ت
ا
ي
ن
ب
1
T
ab
le
7
.
An
a
d
v
er
s
ar
ial
ex
a
m
p
le
g
en
er
ated
wh
en
a
p
p
ly
in
g
Ar
-
Attack
2
Te
x
t
P
r
e
d
i
c
t
i
o
n
ت
ؤا
س
ت
د
ر
ج
م
.
ف
د
ه
و
ةي
ا
ه
ن
ب
.
ةر
ي
ح
و
ةب
ر
غ
ب
ك
ك
ر
ت
ي
ب
ا
ت
ك
0
(
A
b
o
o
k
l
e
a
v
i
n
g
y
o
u
p
e
r
p
l
e
x
e
d
a
n
d
b
a
f
f
l
e
d
.
W
i
t
h
n
o
e
n
d
o
r
p
u
r
p
o
se
.
J
u
st
q
u
e
st
i
o
n
s.)
ب
.
ةر
ي
ح
و
ةب
ر
غ
ب
ك
ك
ر
ت
ي
ب
ا
ت
ك
م
.
ف
د
ه
و
ةي
ا
ه
ن
س
ط
ض
ر
ج
ت
ؤا
1
T
ab
le
8
.
An
a
d
v
er
s
ar
ial
ex
a
m
p
le
g
en
er
ated
wh
en
a
p
p
ly
in
g
Ar
-
Attack
5
Te
x
t
P
r
e
d
i
c
t
i
o
n
ب
ر
ا
ح
ر
س
ا
ي
ار
ك
ش
.
ت
أر
ق
ا
م
لم
ج
أ
ن
م
ت
ا
ق
م
ةع
وم
ج
م
1
(
A
w
o
n
d
e
r
f
u
l
c
o
l
l
e
c
t
i
o
n
o
f
a
r
t
i
c
l
e
s.
T
h
a
n
k
y
o
u
,
Y
a
sser H
a
r
e
b
.
)
ن
م
ط
ا
ق
م
ةع
وم
ج
م
ب
ر
ا
ح
ر
ص
ا
ي
اا
ر
ك
ش
.
طأر
ق
اا
م
لم
ج
أ
0
4
.
2
.
E
v
a
lua
t
i
o
n pro
ce
du
re
B
ef
o
r
e
attac
k
in
g
,
th
e
m
o
d
el
h
ad
an
ac
cu
r
ac
y
o
f
7
4
.
0
8
%.
T
o
ev
alu
ate
th
e
im
p
ac
t
o
f
th
e
attac
k
s
o
n
ac
cu
r
ac
y
,
a
s
u
b
s
et
was
cr
ea
te
d
f
r
o
m
th
e
test
s
et
.
T
h
is
n
ew
s
u
b
s
et
co
n
tain
s
ju
s
t
th
e
in
p
u
ts
th
at
h
av
e
a
c
o
r
r
ec
t
p
r
ed
ictio
n
wh
en
test
in
g
th
e
m
o
d
el
b
ef
o
r
e
attac
k
in
g
.
Usi
n
g
th
e
cr
ea
ted
s
u
b
s
et,
th
e
th
r
e
e
attac
k
s
d
ef
in
e
d
i
n
s
ec
tio
n
3
wer
e
ap
p
lied
.
T
h
en
,
th
e
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
af
ter
ea
ch
attac
k
was
ca
lcu
la
ted
.
I
n
p
u
ts
th
at
ar
e
in
itially
m
is
clas
s
if
ied
ar
e
n
o
t
in
clu
d
ed
,
as
th
ey
co
u
ld
b
ec
o
m
e
co
r
r
ec
tly
class
if
ied
af
ter
th
e
attac
k
,
wh
ich
wo
u
ld
ar
tific
ially
in
cr
ea
s
e
p
o
s
t
-
attac
k
ac
cu
r
ac
y
a
n
d
c
o
n
tr
ad
i
ct
th
e
o
b
jectiv
e
.
4
.
3
.
E
v
a
lua
t
i
o
n o
f
t
he
pro
po
s
ed
a
t
t
a
ck
s
T
ab
le
9
s
h
o
ws
th
e
ev
alu
atio
n
o
f
th
e
p
r
o
p
o
s
ed
attac
k
s
,
Ar
-
At
tack
1
,
Ar
-
Attack
2
,
a
n
d
Ar
-
Attack
5
.
T
h
e
f
ir
s
t
co
lu
m
n
p
r
esen
ts
th
e
m
o
d
el’
s
ac
cu
r
ac
y
in
t
h
e
in
itial
s
ettin
g
s
an
d
u
n
d
er
th
e
s
p
ec
if
ied
a
ttack
s
.
T
h
e
s
ec
o
n
d
co
lu
m
n
p
r
esen
ts
d
ec
r
ea
s
es
o
f
th
e
m
o
d
el’
s
ac
cu
r
ac
y
f
o
r
ea
ch
attac
k
.
T
h
e
f
o
llo
win
g
co
lu
m
n
s
r
ep
o
r
t
th
e
v
alu
es
o
f
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F1
-
s
c
o
r
e,
ea
ch
ac
co
m
p
an
ied
b
y
th
e
ir
r
esp
ec
tiv
e
v
ar
iatio
n
s
to
in
d
i
ca
te
th
e
im
p
ac
t
o
f
ea
ch
attac
k
o
n
th
ese
m
etr
ics.
T
ab
le
9
.
E
x
p
er
im
en
t
’
s
r
esu
lts
A
c
c
u
r
a
c
y
(
%)
D
e
l
t
a
a
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
D
e
l
t
a
p
r
e
c
i
si
o
n
(
%)
R
e
c
a
l
l
(
%)
D
e
l
t
a
r
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
D
e
l
t
a
F
1
-
sco
r
e
(
%)
I
n
i
t
i
a
l
s
e
t
t
i
n
g
s
7
4
.
0
8
-
7
5
.
8
9
-
7
0
.
4
6
-
7
3
.
0
7
-
Ar
-
A
t
t
a
c
k
1
7
1
.
8
4
2
.
2
4
7
2
.
8
6
3
.
0
3
6
9
.
4
5
1
.
0
1
7
1
.
1
1
1
.
9
6
Ar
-
A
t
t
a
c
k
2
7
0
.
8
6
3
.
2
2
7
1
.
5
7
4
.
3
2
6
8
.
9
5
1
.
5
1
7
0
.
2
4
2
.
8
3
Ar
-
A
t
t
a
c
k
5
6
9
.
2
3
4
.
8
5
6
9
.
6
2
6
.
2
7
6
7
.
0
8
3
.
3
8
6
8
.
3
3
4
.
7
4
Fig
u
r
e
4
s
h
o
ws
th
e
r
o
b
u
s
tn
ess
cu
r
v
es
f
o
r
th
e
4
-
ev
alu
atio
n
m
etr
ics:
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e,
u
n
d
er
th
e
in
itial
c
o
n
d
itio
n
a
n
d
th
e
th
r
ee
attac
k
s
ce
n
ar
io
s
(
1
,
2
,
an
d
5
ed
i
ts
)
.
I
t
is
clea
r
th
at
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
941
-
9
5
2
948
ad
v
er
s
ar
ial
ex
am
p
les
ca
n
b
e
s
u
cc
ess
f
u
lly
g
en
er
ate
d
to
f
o
o
l
th
e
m
o
d
el
b
y
m
o
d
if
y
i
n
g
o
r
ad
d
in
g
ju
s
t
o
n
e
ch
ar
ac
ter
to
th
e
m
o
s
t
im
p
o
r
tan
t
wo
r
d
o
f
th
e
in
p
u
t
s
eq
u
e
n
ce
.
T
h
ese
g
en
er
ated
ad
v
er
s
ar
ial
ex
am
p
les
an
d
r
ed
u
ce
d
t
h
e
m
o
d
el
’
s
ac
cu
r
ac
y
f
r
o
m
7
4
.
0
8
to
7
1
.
8
4
%.
W
h
en
th
e
m
o
d
if
icatio
n
is
ap
p
li
ed
to
th
e
two
m
o
s
t
im
p
o
r
tan
t
to
k
en
s
o
f
th
e
in
p
u
t
s
eq
u
en
ce
,
th
e
m
o
d
el’
s
ac
cu
r
ac
y
is
r
ed
u
ce
d
to
7
0
.
8
6
%
a
n
d
it
is
r
ed
u
ce
d
t
o
6
9
.
2
3
%
wh
en
th
e
m
o
d
if
icatio
n
af
f
ec
ts
th
e
f
iv
e
m
o
s
t
im
p
o
r
tan
t
to
k
en
s
.
T
h
e
m
etr
ic
v
alu
es
r
ep
o
r
ted
in
th
e
tab
le
an
d
th
e
r
o
b
u
s
tn
ess
cu
r
v
es
f
o
r
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F1
-
s
co
r
e
ex
h
ib
it
p
atter
n
s
v
e
r
y
s
im
ilar
to
th
o
s
e
o
f
ac
cu
r
ac
y
,
s
u
g
g
esti
n
g
th
at
th
e
t
wo
class
es a
r
e
f
air
ly
b
alan
ce
d
with
n
o
d
o
m
in
an
t
d
is
tr
ib
u
tio
n
b
ias.
Fig
u
r
e
4
.
R
o
b
u
s
tn
ess
cu
r
v
es: m
etr
ics as a
f
u
n
ctio
n
o
f
n
u
m
b
er
o
f
e
d
its
4
.
4
.
Dis
cus
s
io
n
T
ab
les
6
to
8
p
r
o
v
id
e
ex
a
m
p
les
o
f
h
o
w
in
p
u
ts
co
u
ld
b
e
m
is
class
if
ied
wh
en
an
ad
v
er
s
ar
ial
p
er
tu
r
b
atio
n
b
ased
o
n
au
d
ito
r
y
s
im
ilar
ity
is
ap
p
lied
to
th
e
m
o
s
t
im
p
o
r
tan
t
wo
r
d
s
.
T
h
ese
m
in
im
al
ch
an
g
es
ca
n
in
v
o
lv
e
alter
in
g
o
r
in
s
er
tin
g
ju
s
t
o
n
e,
two
,
o
r
f
iv
e
ch
ar
ac
ter
s
in
a
s
en
te
n
ce
.
As
a
r
esu
lt,
a
n
eg
ativ
e
r
ev
iew
ca
n
b
e
tu
r
n
e
d
in
to
a
f
av
o
r
ab
le
o
n
e
o
r
v
ice
v
e
r
s
a.
T
ab
le
9
s
h
o
ws
th
e
ef
f
ec
t
o
f
a
p
p
ly
in
g
th
e
th
r
ee
ad
v
er
s
ar
ial
attac
k
s
to
th
e
d
ec
eiv
ed
m
o
d
el
:
ac
cu
r
ac
y
d
ec
r
ea
s
es
b
y
2
.
2
4
%
f
o
r
Ar
-
A
ttack
1
,
3
.
2
2
%
f
o
r
Ar
-
Attack
2
,
an
d
4
.
8
5
%
f
o
r
Ar
-
Attack
5
with
an
av
er
ag
e
o
f
3
.
4
4
%
f
o
r
th
e
th
r
ee
s
ce
n
ar
io
s
.
T
h
ese
r
esu
lts
s
h
o
w
th
at
B
E
R
T
b
in
ar
y
class
if
ier
s
ar
e
s
en
s
itiv
e
to
co
m
m
o
n
s
p
ellin
g
m
is
tak
es
m
ad
e
b
y
Ar
ab
ic
wr
iter
s
wh
o
m
o
d
if
y
a
ch
a
r
ac
ter
with
a
s
im
ilar
o
n
e
o
r
a
d
d
a
c
h
ar
ac
ter
at
th
e
en
d
o
f
th
e
wo
r
d
.
Fu
r
th
er
m
o
r
e,
g
iv
en
th
at
it
h
as
r
esu
lted
in
a
g
r
ea
ter
r
e
d
u
ctio
n
in
ac
c
u
r
ac
y
,
co
m
b
in
in
g
m
u
ltip
le
m
o
d
if
icatio
n
s
—
ch
an
g
in
g
o
r
ad
d
in
g
two
o
r
f
iv
e
t
o
k
en
s
in
s
tead
o
f
o
n
e
—
lead
s
to
a
m
o
r
e
ef
f
ec
tiv
e
attac
k
.
T
h
is
clea
r
ly
s
h
o
ws
th
e
in
v
er
s
e
r
elatio
n
s
h
ip
b
etwe
en
th
e
d
eg
r
ee
o
f
s
im
ilar
ity
an
d
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
attac
k
.
4
.
5
.
Co
m
plem
ent
a
r
y
a
na
ly
s
i
s
I
n
th
is
s
u
b
s
ec
tio
n
,
co
m
p
lem
e
n
tar
y
an
aly
s
es
ar
e
p
r
esen
ted
to
g
ain
a
d
ee
p
er
u
n
d
er
s
tan
d
i
n
g
o
f
t
h
e
b
eh
av
io
r
an
d
ef
f
ec
tiv
en
ess
o
f
o
u
r
attac
k
.
R
ep
lace
-
o
n
ly
v
er
s
u
s
ad
d
-
o
n
ly
s
tr
ateg
ies
ar
e
f
ir
s
t
ass
es
s
ed
,
an
d
t
h
e
im
p
ac
t
o
f
C
SF
to
k
en
s
elec
tio
n
is
test
ed
v
ia
an
ab
latio
n
.
T
h
e
m
eth
o
d
is
t
h
en
co
m
p
a
r
ed
t
o
a
s
im
p
le
b
aselin
e
an
d
an
e
x
is
tin
g
Ar
ab
ic
attac
k
,
an
d
f
in
ally
th
e
tr
a
n
s
f
er
ab
ilit
y
o
f
th
e
g
en
e
r
ated
ad
v
er
s
ar
ial
ex
am
p
les
is
ev
alu
ated
.
All e
x
p
e
r
im
en
ts
ar
e
co
n
d
u
cted
u
n
d
er
id
e
n
tical
tech
n
ical
s
ettin
g
s
to
en
s
u
r
e
co
m
p
ar
ab
le
r
esu
lts
.
4
.
5
.
1
.
Co
m
po
nent
-
wis
e
a
na
ly
s
is
I
n
th
is
s
u
b
s
ec
tio
n
,
t
h
e
two
at
tack
co
m
p
o
n
en
ts
ar
e
ev
alu
ate
d
s
ep
ar
ately
.
Fo
r
ea
ch
s
am
p
le,
th
e
to
p
to
k
en
is
s
elec
ted
u
s
in
g
C
SF
an
d
ap
p
ly
a
s
in
g
le
p
e
r
tu
r
b
atio
n
,
eith
er
a
r
ep
lace
m
en
t o
r
an
ad
d
itio
n
.
T
h
e
r
esu
ltin
g
ac
cu
r
ac
y
an
d
ac
cu
r
ac
y
d
r
o
p
a
r
e
r
ep
o
r
te
d
in
T
ab
le
1
0
.
T
ab
le
1
0
.
C
o
m
p
o
n
e
n
t
-
wis
e
attac
k
r
esu
lts
A
c
c
u
r
a
c
y
(
%)
D
e
l
t
a
a
c
c
u
r
a
c
y
(
%)
R
e
p
l
a
c
e
o
n
l
y
7
2
.
9
7
1
.
1
1
A
d
d
o
n
l
y
7
2
.
3
5
1
.
7
3
I
t
is
o
b
s
er
v
ed
th
at
ad
d
-
o
n
ly
p
er
f
o
r
m
s
b
etter
th
an
r
ep
lace
-
o
n
ly
.
R
ep
lace
-
o
n
ly
o
n
ly
tar
g
ets
a
s
m
all
s
et
o
f
ch
ar
ac
ter
s
,
s
o
if
th
e
m
o
s
t
im
p
o
r
tan
t
to
k
en
d
o
esn
’
t
co
n
tai
n
an
y
o
f
th
em
,
n
o
ch
a
n
g
e
is
m
ad
e.
B
y
ap
p
ly
i
n
g
r
ep
lace
o
r
ad
d
d
ep
e
n
d
in
g
o
n
t
h
e
to
k
en
,
o
u
r
f
u
ll
attac
k
Ar
-
Attack
1
ac
h
iev
es
b
etter
o
v
er
a
ll
p
er
f
o
r
m
an
ce
th
a
n
eith
er
co
m
p
o
n
en
t a
lo
n
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
d
ve
r
s
a
r
ia
l e
xa
mp
les in
A
r
a
b
ic
la
n
g
u
a
g
e
(
S
a
fa
e
La
a
tya
o
u
i
)
949
4
.
5
.
2
.
Abla
t
io
n
s
t
ud
y
I
n
th
is
s
u
b
s
ec
tio
n
,
th
e
r
o
le
o
f
th
e
C
SF
-
b
ased
to
k
en
s
ele
ctio
n
is
ex
am
in
e
d
b
y
ch
o
o
s
i
n
g
to
k
en
s
ar
b
itra
r
ily
in
s
tead
.
T
h
e
p
e
r
tu
r
b
atio
n
its
elf
is
k
ep
t
th
e
s
am
e,
s
o
d
if
f
er
en
ce
s
in
p
er
f
o
r
m
a
n
ce
d
ir
ec
tly
r
ef
lect
th
e
im
p
ac
t
o
f
t
h
e
C
SF
.
T
h
ese
r
esu
lts
ar
e
co
m
p
ar
e
d
with
th
e
o
r
ig
in
al
C
SF
-
d
r
iv
en
Ar
‑
Attack
1
to
s
ee
h
o
w
m
u
ch
th
e
s
elec
tio
n
m
ec
h
an
is
m
c
o
n
tr
ib
u
t
es
to
th
e
attac
k
’
s
ef
f
ec
tiv
en
es
s
.
Acc
u
r
ac
y
af
ter
attac
k
a
n
d
th
e
d
r
o
p
in
ac
cu
r
ac
y
ar
e
p
r
esen
ted
in
T
ab
le
1
1
.
T
ab
le
1
1
.
A
b
latio
n
s
tu
d
y
r
esu
l
ts
A
c
c
u
r
a
c
y
(
%)
D
e
l
t
a
a
c
c
u
r
a
c
y
(
%)
Ar
-
A
t
t
a
c
k
1
-
o
r
i
g
i
n
a
l
7
1
.
8
4
2
.
2
4
Ar
-
A
t
t
a
c
k
1
-
w
i
t
h
o
u
t
C
S
F
7
2
.
8
8
1
.
2
0
W
ith
C
S
F
-
b
ased
s
elec
tio
n
,
th
e
attac
k
r
ed
u
ce
s
ac
c
u
r
ac
y
b
y
2
.
2
4
%,
wh
ile
u
s
in
g
a
r
an
d
o
m
to
k
en
y
iel
d
o
n
ly
a
1
.
2
0
%
d
r
o
p
.
I
n
o
th
er
wo
r
d
s
,
th
e
C
SF
ac
co
u
n
ts
f
o
r
ab
o
u
t
4
6
%
o
f
th
e
o
b
s
er
v
ed
ac
c
u
r
ac
y
d
e
cr
ea
s
e.
T
h
is
co
n
f
ir
m
s
th
at
tar
g
etin
g
th
e
m
o
s
t im
p
o
r
tan
t to
k
e
n
s
u
b
s
tan
tially
b
o
o
s
ts
th
e
attac
k
’
s
ef
f
ec
tiv
e
n
ess
.
4
.
5
.
3
.
Co
m
pa
riso
n
wit
h r
a
nd
o
m
cha
ra
c
t
er
s
wa
p
I
n
th
is
s
u
b
s
ec
tio
n
,
Ar
-
Attac
k
2
is
co
m
p
a
r
ed
with
a
s
im
p
le
b
aselin
e
th
at
r
an
d
o
m
ly
s
wap
s
two
ch
ar
ac
ter
s
with
in
a
r
an
d
o
m
l
y
s
elec
ted
wo
r
d
.
B
o
th
attac
k
s
ap
p
ly
th
e
s
am
e
n
u
m
b
er
o
f
ed
its
,
m
ak
in
g
th
e
co
m
p
ar
is
o
n
m
ea
n
in
g
f
u
l.
Acc
u
r
ac
y
af
ter
attac
k
an
d
t
h
e
d
r
o
p
i
n
ac
cu
r
ac
y
ar
e
s
h
o
wn
in
T
ab
le
1
2
.
T
ab
le
1
2
.
C
o
m
p
ar
ativ
e
r
esu
lts
with
r
an
d
o
m
ch
ar
ac
ter
s
wap
A
c
c
u
r
a
c
y
(
%)
D
e
l
t
a
a
c
c
u
r
a
c
y
(
%)
Ar
-
A
t
t
a
c
k
2
7
0
.
8
6
3
.
2
2
C
h
a
r
a
c
t
e
r
sw
a
p
b
a
s
e
l
i
n
e
7
2
.
3
4
1
.
7
4
T
h
e
ch
ar
ac
ter
s
wap
b
aselin
e
y
ield
s
an
ac
cu
r
ac
y
d
r
o
p
o
f
1
.
7
4
%,
wh
er
ea
s
o
u
r
A
r
-
Attack
2
ac
h
iev
es
a
lar
g
er
d
ec
r
ea
s
e
o
f
3
.
2
2
%
u
n
d
er
th
e
s
am
e
e
d
it
d
is
tan
ce
(
two
ed
its
)
.
E
v
en
Ar
-
Attack
1
,
w
h
ich
u
s
es
o
n
l
y
o
n
e
ed
it,
s
u
r
p
ass
es
th
e
b
aselin
e
w
ith
a
2
.
2
4
%
d
r
o
p
.
T
h
ese
r
esu
l
ts
s
h
o
w
th
at
o
u
r
p
r
o
p
o
s
ed
attac
k
s
g
en
er
ate
m
o
r
e
ef
f
ec
tiv
e
p
er
tu
r
b
atio
n
s
th
a
n
s
im
p
le
ch
ar
ac
ter
s
wap
s
,
ev
e
n
wi
th
f
ewe
r
ed
its
.
4
.
5
.
4
.
Co
m
pa
riso
n wit
h
ex
is
t
i
ng
Ara
bic
a
t
t
a
c
k
I
n
th
is
s
u
b
s
ec
tio
n
,
Ar
-
Attack
1
is
co
m
p
ar
ed
with
th
e
Ar
a
b
ic
attac
k
in
tr
o
d
u
ce
d
in
[
3
4
]
.
B
o
th
m
eth
o
d
s
ar
e
ev
alu
ated
u
n
d
er
th
e
s
am
e
n
u
m
b
er
o
f
ed
its
to
k
ee
p
th
e
co
m
p
ar
is
o
n
f
air
.
W
h
ile
o
u
r
ap
p
r
o
ac
h
r
elies
o
n
au
d
ito
r
y
s
im
ilar
ity
b
etwe
en
Ar
ab
ic
ch
ar
ac
ter
s
,
th
eir
s
is
b
ase
d
o
n
v
is
u
al
s
im
ilar
ity
.
A
q
u
an
t
itativ
e
ev
alu
atio
n
is
p
er
f
o
r
m
ed
b
ased
o
n
attac
k
ef
f
ec
tiv
en
ess
,
an
d
a
q
u
alitativ
e
o
n
e
ass
ess
in
g
h
u
m
an
p
er
ce
p
tib
i
lity
.
i)
A
t
t
a
c
k
ef
f
ec
t
i
v
en
e
s
s
e
v
a
l
u
a
ti
o
n
:
th
e
im
p
ac
t
o
f
ea
c
h
a
t
ta
c
k
o
n
t
h
e
c
l
a
s
s
i
f
i
e
r
’
s
a
c
cu
r
a
c
y
i
s
f
ir
s
t
e
v
a
l
u
a
t
e
d
.
R
e
s
u
l
t
s
a
r
e
s
u
m
m
a
r
i
z
e
d
i
n
T
a
b
l
e
1
3
.
T
h
e
f
ir
s
t
c
o
l
u
m
n
r
e
p
o
r
t
s
ac
c
u
r
a
c
y
a
f
te
r
t
h
e
a
t
t
a
ck
,
a
n
d
t
h
e
s
e
c
o
n
d
c
o
lu
m
n
s
h
o
ws
t
h
e
d
r
o
p
i
n
a
c
cu
r
a
cy
r
e
l
a
t
iv
e
t
o
t
h
e
o
r
i
g
in
a
l
cl
e
a
n
m
o
d
e
l.
A
s
s
h
o
wn
in
T
a
b
l
e
1
3
,
th
e
v
i
s
u
a
l
-
s
i
m
i
l
a
r
i
t
y
m
e
th
o
d
p
r
o
d
u
c
e
s
a
la
r
g
e
r
a
c
cu
r
a
c
y
d
r
o
p
o
n
th
e
t
ar
g
e
t
c
l
a
s
s
i
f
i
er
t
h
a
n
Ar
-
A
t
t
ac
k
1
w
h
e
n
b
o
th
m
e
t
h
o
d
s
u
s
e
t
h
e
s
a
m
e
n
u
m
b
e
r
o
f
ed
i
t
s
.
T
h
e
s
e
r
e
s
u
l
t
s
in
d
i
c
a
t
e
t
h
a
t
t
h
e
v
i
s
u
a
l
-
b
a
s
e
d
m
e
th
o
d
y
i
e
l
d
s
a
l
a
r
g
er
p
r
o
p
o
r
t
i
o
n
o
f
in
p
u
t
s
t
h
a
t
t
h
e
c
l
a
s
s
i
f
i
e
r
m
i
s
l
a
b
e
l
s
t
h
an
A
r
-
A
t
t
a
c
k
1
.
T
ab
le
1
3
.
Attack
e
f
f
ec
tiv
en
ess
co
m
p
ar
is
o
n
A
c
c
u
r
a
c
y
(
%)
D
e
l
t
a
A
c
c
u
r
a
c
y
(
%)
Ar
-
A
t
t
a
c
k
1
(
a
u
d
i
t
o
r
y
-
b
a
s
e
d
,
o
u
r
s)
7
1
.
8
4
2
.
2
4
V
i
su
a
l
-
b
a
se
d
A
r
a
b
i
c
a
t
t
a
c
k
6
9
.
8
4
.
2
8
ii)
Hu
m
an
p
er
ce
p
tib
ilit
y
ev
alu
at
io
n
:
to
co
m
p
lem
e
n
t
th
e
q
u
a
n
titativ
e
r
esu
lts
,
a
h
u
m
an
ev
alu
atio
n
was
co
n
d
u
cte
d
with
2
0
n
ativ
e
an
n
o
tato
r
s
o
n
1
0
0
p
er
tu
r
b
ed
r
e
v
iews,
5
0
v
is
u
al
-
b
ased
an
d
th
ei
r
5
0
au
d
ito
r
y
co
u
n
ter
p
a
r
ts
.
E
ac
h
an
n
o
tato
r
r
ec
eiv
ed
a
d
is
tin
ct
test
s
et,
e
n
s
u
r
in
g
n
o
o
n
e
s
aw
th
e
s
am
e
r
ev
iew
u
n
d
er
b
o
th
p
e
r
tu
r
b
atio
n
ty
p
es.
E
ac
h
tr
ial
last
ed
1
5
s
ec
o
n
d
s
.
R
esu
lts
ar
e
r
ep
o
r
te
d
in
T
ab
le
1
4
.
T
h
e
r
esu
lts
s
h
o
w
th
at
p
er
t
u
r
b
atio
n
s
f
r
o
m
Ar
-
Attack
1
wer
e
less
o
f
ten
d
etec
ted
th
an
t
h
o
s
e
f
r
o
m
th
e
v
is
u
al
-
b
ased
attac
k
.
T
h
is
in
d
icate
s
th
at
au
d
ito
r
y
s
u
b
s
titu
ti
o
n
s
ap
p
ea
r
m
o
r
e
n
atu
r
al
an
d
h
ar
d
er
to
n
o
tice
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
941
-
9
5
2
950
W
ith
in
Ar
-
Attack
1
,
6
9
%
o
f
a
u
d
ito
r
y
-
ad
d
a
n
d
3
3
%
o
f
au
d
it
o
r
y
-
r
e
p
lace
ca
s
es
wen
t
u
n
n
o
ti
ce
d
,
s
u
g
g
esti
n
g
th
at
th
e
ad
d
itiv
e
m
o
d
if
icatio
n
is
le
s
s
p
er
ce
p
tib
le
th
an
r
e
p
lace
m
e
n
t o
n
es.
T
ab
le
1
4
.
H
u
m
an
p
er
ce
p
tib
ilit
y
co
m
p
a
r
is
o
n
N
u
mb
e
r
o
f
p
e
r
t
u
r
b
e
d
r
e
v
i
e
w
s
D
e
t
e
c
t
i
o
n
ra
te
(
%)
Ar
-
A
t
t
a
c
k
1
(
a
u
d
i
t
o
r
y
-
b
a
s
e
d
,
o
u
r
s)
50
44
V
i
su
a
l
-
b
a
se
d
A
r
a
b
i
c
a
t
t
a
c
k
50
52
4
.
5
.
5
.
T
ra
ns
f
er
a
bil
it
y
a
na
ly
s
is
Ho
w
ad
v
er
s
ar
ial
ex
am
p
les
g
en
er
ated
u
s
in
g
t
h
e
Ar
-
Attack
1
m
et
h
o
d
o
n
th
e
s
o
u
r
ce
B
E
R
T
-
b
ased
m
o
d
el
tr
an
s
f
er
to
Ar
aBER
T
—
n
am
ely
ar
ab
er
t
-
s
en
tim
en
t
-
m
o
d
el
-
Mu
h
an
n
e
d
Sh
,
a
p
r
etr
a
in
ed
Ar
ab
ic
m
o
d
el
f
in
e
-
tu
n
e
d
f
o
r
b
i
n
ar
y
SA
a
v
ailab
le
o
n
H
u
g
g
in
g
Face
—
is
ev
alu
ated
.
T
h
e
tr
an
s
f
er
ab
ilit
y
is
ass
es
s
ed
u
s
in
g
th
e
tr
an
s
f
er
s
u
cc
ess
r
ate
(
T
SR
)
d
e
f
in
ed
as
th
e
r
atio
o
f
s
u
cc
ess
f
u
l
ad
v
er
s
ar
ial
attac
k
s
o
n
th
e
tar
g
et
m
o
d
el
to
t
h
o
s
e
th
at
s
u
cc
ee
d
o
n
th
e
s
o
u
r
ce
m
o
d
el.
R
esu
lts
ar
e
r
ep
o
r
ted
in
T
ab
le
1
5
.
A
b
o
u
t
3
1
.
4
7
%
o
f
th
e
ad
v
e
r
s
ar
ial
ex
am
p
les
th
at
f
o
o
led
B
E
R
T
also
f
o
o
led
Ar
aBER
T
,
s
h
o
win
g
m
o
d
er
ate
tr
a
n
s
f
er
ab
ilit
y
.
T
ab
le
1
5
.
T
r
an
s
f
er
a
b
ilit
y
r
esu
lts
f
o
r
Ar
aBER
T
N
u
mb
e
r
o
f
su
c
c
e
sses
o
n
s
o
u
r
c
e
N
u
mb
e
r
o
f
su
c
c
e
sses
o
n
t
a
r
g
e
t
TSR
(
%)
A
r
a
B
ER
T
6
9
6
2
1
9
3
1
.
4
7
5.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
p
r
o
p
o
s
ed
a
c
h
ar
a
cter
-
lev
el
m
eth
o
d
to
g
en
er
at
e
ad
v
er
s
ar
ial
ex
am
p
les
in
A
r
ab
ic.
T
h
e
s
u
g
g
ested
ap
p
r
o
ac
h
is
s
im
p
le
an
d
h
ar
d
l
y
p
er
ce
p
tib
le,
r
e
q
u
ir
i
n
g
o
n
ly
a
f
ew
ca
r
ef
u
lly
s
elec
ted
tr
an
s
f
o
r
m
atio
n
s
.
I
t
o
p
er
ates
u
n
d
er
BB
s
ettin
g
s
,
with
o
u
t
ac
ce
s
s
to
th
e
in
ter
n
al
p
ar
am
eter
s
o
f
th
e
m
o
d
el.
T
h
e
g
en
er
ated
ad
v
er
s
ar
ial
ex
am
p
les
ap
p
ea
r
n
atu
r
al
y
et
ef
f
ec
tiv
ely
co
n
f
u
s
e
th
e
class
if
ier
,
lead
in
g
to
an
av
er
ag
e
ac
cu
r
ac
y
d
r
o
p
o
f
3
.
4
4
%
f
o
r
a
B
E
R
T
m
o
d
el
r
etr
ain
ed
o
n
th
e
B
R
AD
Ar
ab
ic
d
ataset
,
th
u
s
d
em
o
n
s
tr
atin
g
th
e
m
eth
o
d
’
s
ef
f
ec
tiv
en
ess
.
C
o
m
p
ar
ed
to
a
ch
ar
ac
ter
-
s
wap
b
aselin
e,
o
u
r
m
eth
o
d
ca
u
s
es
a
lar
g
er
ac
c
u
r
a
cy
d
r
o
p
.
W
h
ile
th
e
v
is
u
al
-
b
ased
r
ep
lace
m
eth
o
d
ac
h
iev
es
a
lar
g
er
ac
cu
r
ac
y
d
r
o
p
o
n
Ar
ab
ic
tex
t,
o
u
r
au
d
ito
r
y
-
ad
d
p
e
r
tu
r
b
atio
n
s
d
em
o
n
s
tr
ate
s
u
p
er
io
r
h
u
m
an
i
m
p
er
ce
p
tib
ilit
y
.
B
u
ild
in
g
o
n
th
ese
r
esu
lts
,
f
u
tu
r
e
wo
r
k
co
u
l
d
ex
p
lo
r
e
ad
d
itio
n
al
attac
k
v
ar
ian
ts
,
ex
ten
d
th
e
m
eth
o
d
to
au
d
io
-
b
ased
in
p
u
ts
,
an
d
ass
es
s
ad
v
er
s
ar
ial
r
o
b
u
s
tn
ess
b
ey
o
n
d
ch
ar
ac
ter
-
lev
el
p
er
tu
r
b
atio
n
s
.
ACK
NO
WL
E
DG
M
E
N
T
S
W
e
wo
u
ld
lik
e
to
ex
p
r
ess
o
u
r
g
r
atitu
d
e
to
Dr
.
Mo
s
taf
a
Ay
y
a
d
f
o
r
h
is
E
n
g
lis
h
r
ev
iew
o
f
th
i
s
p
ap
er
.
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
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Saf
ae
L
aa
ty
ao
u
i
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Mo
h
am
m
ed
Sab
e
r
✓
✓
✓
✓
✓
✓
✓
✓
✓
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
Evaluation Warning : The document was created with Spire.PDF for Python.