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
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
,
p
p
.
4
6
8
4
~
4
6
9
3
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
6
.
p
p
4
6
8
4
-
4
6
9
3
4684
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Fine
-
tuning mu
l
ti
ling
ua
l t
ra
nsfo
rm
ers for
H
ing
lish s
entimen
t
a
na
ly
sis
:
a
com
pa
ra
tive eva
lua
tion
with
BiL
STM
J
y
o
t
i S
.
Ver
m
a
,
J
a
im
in N
.
U
nd
a
v
ia
S
mt
.
C
h
a
n
d
a
b
e
n
M
o
h
a
n
b
h
a
i
P
a
t
e
l
I
n
s
t
i
t
u
t
e
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s
,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s,
C
h
a
r
o
t
a
r
U
n
i
v
e
r
si
t
y
o
f
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
A
n
a
n
d
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
r
5
,
2
0
2
5
R
ev
is
ed
Au
g
2
5
,
2
0
2
5
Acc
ep
ted
Sep
7
,
2
0
2
5
G
ro
win
g
tren
d
o
f
c
o
d
e
-
m
ix
in
g
in
lan
g
u
a
g
e
s,
in
t
h
e
f
o
rm
o
f
Hin
g
li
sh
,
g
re
a
tl
y
tes
ts
th
e
sk
il
ls
o
f
c
o
n
v
e
n
t
io
n
a
l
se
n
ti
m
e
n
t
a
n
a
l
y
sis
t
o
o
ls.
Th
e
re
se
a
rc
h
c
o
n
tri
b
u
tes
a
fin
e
-
t
u
n
e
d
m
u
lt
il
in
g
u
a
l
tran
sf
o
rm
e
r
m
o
d
e
l
b
u
il
t
e
x
c
lu
siv
e
ly
f
o
r
c
las
sify
in
g
se
n
ti
m
e
n
t
o
f
Hin
g
li
s
h
c
u
st
o
m
e
r
re
v
iew
s.
Dra
win
g
f
ro
m
p
re
-
train
e
d
BERT
-
b
a
se
-
m
u
lt
il
in
g
u
a
l
-
c
a
se
a
r
c
h
it
e
c
tu
re
,
th
e
m
o
d
e
l
g
e
ts
tran
sfo
rm
e
d
with
t
h
e
p
r
o
c
e
ss
o
f
fin
e
-
t
u
n
i
n
g
t
h
e
sa
m
e
o
n
sy
n
th
e
ti
c
a
ll
y
p
re
p
a
re
d
a
n
d
b
a
lan
c
e
d
d
a
tas
e
t
sim
u
latin
g
p
o
siti
v
e
,
n
e
g
a
ti
v
e
,
a
n
d
n
e
u
tral
se
n
ti
m
e
n
t
s.
S
o
p
h
isti
c
a
ted
m
e
th
o
d
s
li
k
e
f
o
c
a
l
l
o
ss
fo
r
a
d
d
re
ss
in
g
th
e
c
las
s
imb
a
lan
c
e
a
n
d
m
i
x
e
d
p
re
c
isio
n
t
ra
in
in
g
fo
r
m
a
x
imiz
a
ti
o
n
o
f
c
o
m
p
u
tati
o
n
a
l
e
ffe
c
ti
v
e
n
e
ss
a
re
e
m
b
e
d
d
e
d
with
i
n
th
e
train
i
n
g
p
r
o
c
e
ss
.
Ex
p
e
rime
n
tal
re
su
lt
s
su
g
g
e
st
t
h
a
t
th
e
p
r
o
p
o
se
d
m
e
th
o
d
sig
n
ifi
c
a
n
tl
y
c
a
p
tu
re
s
t
h
e
fin
e
-
g
ra
in
e
d
li
n
g
u
isti
c
p
a
tt
e
rn
s
o
f
c
o
d
e
-
m
ix
e
d
tex
t,
imp
r
o
v
in
g
se
n
ti
m
e
n
t
c
las
sifica
ti
o
n
a
c
c
u
ra
c
y
.
Th
e
re
su
lt
s
sh
o
w
p
ro
m
isin
g
p
o
te
n
ti
a
l
f
o
r
e
n
h
a
n
c
i
n
g
c
u
sto
m
e
r
fe
e
d
b
a
c
k
a
n
a
ly
sis
i
n
e
-
c
o
m
m
e
rc
e
,
so
c
ial
m
e
d
ia
m
o
n
it
o
r
in
g
,
a
n
d
c
u
sto
m
e
r
su
p
p
o
rt
u
se
c
a
se
s,
wh
e
re
it
is
c
ru
c
ial
to
c
o
m
p
re
h
e
n
d
t
h
e
se
n
ti
m
e
n
t
b
e
h
i
n
d
c
o
d
e
-
m
ix
e
d
re
v
iew
s.
K
ey
w
o
r
d
s
:
B
o
o
s
tin
g
alg
o
r
ith
m
Hin
d
i
-
E
n
g
lis
h
r
atin
g
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
Sen
tim
en
t a
n
aly
s
is
Stack
in
g
en
s
em
b
le
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
:
J
y
o
ti S
.
Ver
m
a
Sm
t.
C
h
an
d
ab
en
Mo
h
an
b
h
ai
Patel
I
n
s
titu
te
o
f
C
o
m
p
u
ter
A
p
p
licatio
n
s
Facu
lty
o
f
C
o
m
p
u
ter
Scien
ce
an
d
Ap
p
licatio
n
s
,
C
h
ar
o
tar
Un
iv
er
s
ity
o
f
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
C
HA
R
USAT
C
am
p
u
s
,
C
h
an
g
a,
An
an
d
,
G
u
jar
at
3
8
8
4
2
1
,
I
n
d
ia
E
m
ail: jy
o
ti.s.v
er
m
aa
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
As
co
m
p
an
ies
r
ap
id
ly
d
ig
itize
an
d
co
n
s
u
m
er
s
in
c
r
ea
s
in
g
ly
r
ely
o
n
u
s
er
-
g
en
er
ate
d
c
o
n
ten
t,
cu
s
to
m
er
r
ev
iews
ar
e
n
o
w
a
d
ec
is
iv
e
f
ac
to
r
in
p
u
r
c
h
ase
d
ec
is
io
n
s
.
Sen
tim
en
t
an
aly
s
is
,
a
f
ield
o
f
n
atu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
NL
P),
allo
ws
co
m
p
an
ies
to
g
ain
m
ea
n
in
g
f
u
l
in
s
ig
h
ts
f
r
o
m
r
ev
iews
b
y
ca
teg
o
r
izin
g
th
em
as
p
o
s
itiv
e,
n
eg
ativ
e,
o
r
n
eu
tr
al.
I
n
I
n
d
ia
,
m
o
s
t
o
f
th
e
u
s
er
s
r
ep
r
esen
t
th
e
r
ev
iews
u
s
in
g
t
h
e
E
n
g
lis
h
lan
g
u
ag
e.
B
u
t
to
ef
f
ec
tiv
ely
r
ep
r
esen
t
th
eir
th
o
u
g
h
ts
,
th
ey
u
s
e
E
n
g
li
s
h
to
wr
ite
r
ev
iews
,
b
u
t
in
Hin
d
i
s
p
ee
ch
,
lik
e
“
Mu
jh
e
d
r
ess
ma
te
r
ia
l
b
a
h
o
t
p
a
s
a
n
d
a
a
ya
(
I
r
ea
lly
lik
e
d
th
e
d
r
ess
m
ater
ial
)
”
o
r
“
E
k
d
u
m
b
ek
a
r
crea
m
h
a
i
ma
t
len
a
(
T
h
is
c
r
ea
m
is
a
b
s
o
lu
tely
ter
r
ib
le,
d
o
n
'
t
b
u
y
it
)
”.
Su
ch
a
p
o
p
u
lar
co
d
e
-
m
ix
ed
lan
g
u
ag
e
is
Hin
g
lis
h
,
a
m
ix
tu
r
e
o
f
E
n
g
lis
h
an
d
Hin
d
i,
u
s
u
ally
wr
itten
.
Hin
g
lis
h
c
an
b
e
r
eg
u
lar
ly
n
o
ticed
in
e
-
co
m
m
er
ce
r
ev
iews,
s
o
cial
m
ed
ia
p
o
s
tin
g
s
,
an
d
c
u
s
to
m
er
f
ee
d
b
ac
k
o
n
t
h
e
in
ter
n
et,
wh
ich
is
a
p
r
ec
io
u
s
b
u
t
d
if
f
icu
lt
s
o
u
r
ce
f
o
r
s
en
tim
en
t
an
aly
s
is
.
Yet,
ex
is
tin
g
s
en
tim
en
t
an
aly
s
is
m
o
d
els
m
ain
ly
o
p
er
ate
o
n
m
o
n
o
lin
g
u
al
tex
ts
an
d
th
er
ef
o
r
e
ca
n
n
o
t
b
e
ef
f
ec
tiv
e
i
n
m
ar
k
ets wh
er
e
co
n
s
u
m
er
s
u
s
e
co
d
e
-
m
ix
e
d
lan
g
u
ag
es c
o
m
m
o
n
ly
.
T
h
e
ab
s
en
ce
o
f
u
n
if
ied
s
p
ellin
g
p
r
ac
tices,
ter
m
in
o
lo
g
y
,
an
d
m
ix
e
d
g
r
am
m
atica
l
s
tr
u
ctu
r
es
p
r
esen
ts
a
m
ajo
r
p
r
o
b
lem
f
o
r
co
n
v
en
tio
n
al
NL
P
m
o
d
els.
Als
o
,
th
e
p
r
e
-
tr
ain
ed
s
en
tim
en
t
a
n
aly
s
is
m
o
d
els
ar
e
u
s
u
ally
tr
ain
ed
f
o
r
E
n
g
lis
h
o
r
Hin
d
i
in
d
iv
id
u
ally
,
d
o
n
o
t
p
r
o
p
er
ly
u
n
d
e
r
s
tan
d
Hin
g
lis
h
s
em
an
tics
an
d
co
n
tex
t
,
a
n
d
m
i
s
class
if
y
s
en
tim
en
t.
T
o
co
u
n
ter
ac
t
t
h
ese
is
s
u
es,
o
u
r
s
tu
d
y
estab
lis
h
es
a
s
en
tim
en
t
class
if
icatio
n
.
T
h
e
m
o
d
el
is
p
ar
ticu
lar
ly
g
ea
r
ed
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
F
in
e
-
tu
n
in
g
mu
ltil
in
g
u
a
l tra
n
s
fo
r
mers
fo
r
H
in
g
lis
h
s
en
timen
t a
n
a
lysi
s
:
a
…
(
Jy
o
ti
S
.
V
erma
)
4685
f
o
r
Hin
g
lis
h
cu
s
to
m
er
c
o
m
m
en
ts
.
W
e
em
p
lo
y
th
e
ca
p
ab
il
ity
o
f
p
r
e
-
t
r
ain
ed
m
u
ltil
in
g
u
al
tr
an
s
f
o
r
m
er
s
—
a
B
E
R
T
-
b
ase
-
m
u
ltil
in
g
u
al
-
ca
s
ed
m
o
d
el
s
p
ec
if
ically
to
co
m
b
at
th
e
in
tr
icac
ies
s
u
r
r
o
u
n
d
in
g
co
d
e
-
m
ix
la
n
g
u
a
g
e
d
ata.
T
h
e
m
o
d
el
is
th
en
f
in
e
-
t
u
n
ed
o
n
an
e
q
u
itab
ly
s
p
lit,
s
y
n
th
etic
co
r
p
u
s
o
f
Hin
g
lis
h
d
at
a
in
wh
ich
ea
ch
o
f
th
e
th
r
ee
class
es
o
f
s
en
tim
en
t
(
p
o
s
itiv
e,
n
eg
ativ
e,
an
d
n
eu
tr
a
l)
h
as
an
eq
u
iv
alen
t
p
r
esen
ce
.
T
o
f
u
r
th
e
r
en
h
an
ce
p
er
f
o
r
m
an
ce
,
we
h
av
e
also
e
m
p
lo
y
ed
cu
ttin
g
-
e
d
g
e
m
eth
o
d
s
lik
e
f
o
ca
l
lo
s
s
to
c
o
u
n
ter
cl
ass
im
b
alan
ce
an
d
m
ix
ed
p
r
ec
is
io
n
tr
ain
in
g
t
o
m
a
x
im
ize
r
eso
u
r
ce
e
f
f
icien
cy
.
Ou
r
ap
p
r
o
a
ch
i
n
c
lu
d
es d
e
tai
le
d
d
ata
p
r
e
p
r
o
c
ess
i
n
g
(
t
o
k
e
n
iz
a
tio
n
,
p
a
d
d
i
n
g
,
a
n
d
tr
u
n
c
ati
o
n
)
,
s
y
s
te
m
a
tic
d
at
a
s
p
lit
i
n
t
o
t
r
ai
n
i
n
g
a
n
d
v
a
li
d
at
io
n
s
e
ts
,
a
n
d
a
r
o
b
u
s
t
f
in
e
-
t
u
n
i
n
g
p
r
o
ce
s
s
.
E
x
p
er
im
en
tal
r
esu
l
ts
s
h
o
w
t
h
a
t
th
e
p
r
o
p
o
s
e
d
s
o
lu
ti
o
n
s
i
g
n
i
f
ic
a
n
tl
y
im
p
r
o
v
es
s
e
n
ti
m
e
n
t
c
lass
i
f
i
c
ati
o
n
ac
c
u
r
a
cy
o
n
Hi
n
g
lis
h
te
x
t
.
T
h
e
c
o
n
t
r
i
b
u
ti
o
n
o
f
th
is
wo
r
k
h
as
d
ee
p
i
m
p
li
c
ati
o
n
s
f
o
r
ap
p
l
ic
ati
o
n
in
s
o
ci
al
m
ed
ia
m
o
n
i
to
r
i
n
g
,
e
-
co
m
m
er
ce
,
a
n
d
cu
s
to
m
e
r
s
u
p
p
o
r
t s
y
s
t
em
s
,
w
h
e
r
e
s
e
n
ti
m
en
t
a
n
al
y
s
is
o
f
s
o
p
h
is
ti
ca
te
d
,
c
o
d
e
-
m
i
x
ed
f
ee
d
b
a
c
k
is
o
f
c
r
i
ti
ca
l
i
m
p
o
r
ta
n
ce
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
Sen
tim
en
t
an
aly
s
is
h
as
e
m
e
r
g
ed
as
o
n
e
o
f
t
h
e
m
o
s
t
v
alu
ab
le
ap
p
licatio
n
s
f
o
r
u
n
d
er
s
tan
d
in
g
co
n
s
u
m
er
o
p
in
io
n
s
ac
r
o
s
s
d
ig
ital
p
latf
o
r
m
s
.
E
ar
ly
a
p
p
r
o
ac
h
e
s
p
r
im
ar
ily
u
tili
ze
d
m
ac
h
i
n
e
l
ea
r
n
in
g
tech
n
iq
u
es
s
u
ch
as
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
(
SVM
)
a
n
d
r
an
d
o
m
f
o
r
est
co
m
b
i
n
ed
with
lex
ico
n
-
b
ase
d
m
eth
o
d
s
[
1
]
,
[
2
]
.
Ho
wev
er
,
th
ese
m
eth
o
d
s
d
e
m
o
n
s
tr
ated
s
ig
n
if
ican
t
lim
itatio
n
s
in
h
an
d
lin
g
im
p
licit
s
en
tim
en
t,
co
d
e
-
m
ix
e
d
tex
ts
,
an
d
d
o
m
ain
ad
ap
tatio
n
ch
allen
g
es
[
1
]
,
[
2
]
.
T
h
e
f
ield
ad
v
an
ce
d
co
n
s
id
er
a
b
ly
with
th
e
em
er
g
en
ce
o
f
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es.
Mo
d
els
in
co
r
p
o
r
atin
g
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN
s
)
,
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
,
an
d
h
y
b
r
id
f
r
am
ew
o
r
k
s
s
h
o
wed
im
p
r
o
v
ed
p
er
f
o
r
m
an
ce
b
y
lear
n
in
g
s
eq
u
en
tial
p
atter
n
s
f
r
o
m
tex
t
[
3
]
,
[
4
]
.
Desp
ite
th
ese
im
p
r
o
v
em
en
ts
,
th
e
u
n
s
tr
u
ctu
r
e
d
n
atu
r
e
o
f
s
o
cial
m
ed
ia
co
n
ten
t
co
n
tin
u
ed
to
d
em
an
d
m
o
r
e
s
ca
lab
le
s
o
lu
tio
n
s
[
5
]
,
[
6
]
.
A
b
r
ea
k
th
r
o
u
g
h
ca
m
e
with
atten
tio
n
-
en
h
a
n
ce
d
B
i
-
L
STM
ar
ch
itectu
r
es
th
at
co
u
ld
ca
p
tu
r
e
ess
en
tial
f
ea
tu
r
es
in
r
aw
te
x
t
in
p
u
ts
[
7
]
,
[
8
]
.
Yet,
co
d
e
-
m
ix
e
d
la
n
g
u
ag
es
lik
e
Hin
g
lis
h
r
em
ain
e
d
p
r
o
b
lem
atic
d
u
e
t
o
th
eir
p
h
o
n
etic
v
ar
iatio
n
s
an
d
i
r
r
eg
u
lar
s
y
n
tax
[
9
]
,
[
1
0
]
.
R
ec
en
t
m
u
ltil
in
g
u
al
m
o
d
els
lik
e
Mu
R
I
L
,
XL
M
-
R
,
an
d
I
n
d
icB
E
R
T
h
av
e
s
h
o
wn
p
ar
ti
cu
lar
p
r
o
m
is
e
f
o
r
I
n
d
ian
lan
g
u
ag
es
[
1
1
]
–
[
1
3
]
.
W
h
ile
o
u
tp
e
r
f
o
r
m
in
g
p
r
ev
io
u
s
ap
p
r
o
ac
h
es,
th
ey
s
till
s
tr
u
g
g
l
e
with
s
ar
ca
s
m
d
etec
tio
n
an
d
r
ea
l
-
tim
e
ef
f
icien
cy
[
1
4
]
,
[
1
5
]
.
T
h
e
r
esear
ch
lan
d
s
ca
p
e
f
o
r
E
n
g
lis
h
an
d
H
in
d
i
s
en
tim
en
t
an
aly
s
is
is
we
ll
-
estab
lis
h
ed
[
1
4
]
,
[
1
6
]
.
Ho
wev
er
,
co
d
e
-
m
ix
ed
f
o
r
m
s
lik
e
Hin
g
lis
h
p
r
esen
t
u
n
iq
u
e
c
h
allen
g
es
[
1
7
]
,
[
1
8
]
.
Mo
d
els
tr
ain
ed
o
n
p
u
r
e
lan
g
u
ag
es
o
f
ten
f
ail
to
g
en
er
alize
d
u
e
to
s
y
n
tactic
v
ar
iatio
n
[
1
9
]
,
[
2
0
]
.
I
n
itial
att
em
p
ts
u
s
in
g
r
u
le
-
b
ased
[
2
1
]
an
d
lex
ico
n
-
b
ased
m
eth
o
d
s
[
2
2
]
f
ac
ed
lim
itatio
n
s
f
r
o
m
lin
g
u
is
tic
am
b
ig
u
ity
.
T
h
e
Sem
E
v
al
-
2
0
2
0
T
ask
9
b
en
ch
m
ar
k
s
ig
n
if
ica
n
tly
ad
v
an
ce
d
th
e
f
ield
[
2
1
]
,
[
2
3
]
,
wh
ile
ex
p
an
d
ed
d
atasets
f
r
o
m
s
o
cial
p
latf
o
r
m
s
e
n
ab
led
b
ette
r
tr
ain
in
g
[
1
1
]
.
Key
ch
allen
g
es
in
Hin
g
lis
h
a
n
aly
s
is
in
clu
d
e
v
o
ca
b
u
lar
y
s
tan
d
ar
d
izatio
n
a
n
d
p
h
o
n
etic
v
ar
i
atio
n
s
[
1
8
]
,
[
2
4
]
.
T
r
an
s
f
o
r
m
e
r
m
o
d
els
lik
e
B
E
R
T
[
2
5
]
,
XL
M
-
R
[
1
2
]
,
a
n
d
Mu
R
I
L
[
1
1
]
h
a
v
e
s
h
o
wn
p
r
o
m
is
e
wh
e
n
p
r
o
p
er
l
y
f
in
e
-
tu
n
e
d
.
Hy
b
r
id
ap
p
r
o
ac
h
e
s
co
m
b
in
in
g
lex
ic
o
n
s
with
b
i
d
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
iLST
M
)
ar
ch
itectu
r
es
h
av
e
d
em
o
n
s
tr
ated
s
u
p
er
io
r
p
er
f
o
r
m
a
n
ce
.
Fas
tTe
x
t
em
b
ed
d
in
g
s
p
r
o
v
e
p
ar
tic
u
lar
ly
ef
f
ec
tiv
e
f
o
r
ca
p
tu
r
in
g
p
h
o
n
etic
v
ar
iatio
n
s
[
3
]
.
Desp
ite
p
r
o
g
r
ess
,
r
ea
l
-
tim
e
d
ep
lo
y
m
e
n
t
r
em
a
in
s
c
h
allen
g
in
g
.
Fu
tu
r
e
d
ir
ec
tio
n
s
in
clu
d
e
c
o
r
p
u
s
ex
p
a
n
s
io
n
,
co
n
tr
asti
v
e
lear
n
in
g
,
an
d
o
p
tim
ized
tr
a
n
s
f
o
r
m
e
r
d
ep
l
o
y
m
en
t
[
2
3
]
.
3.
B
E
R
T
A
L
G
O
RIT
H
M
B
E
R
T
is
b
u
ilt
u
p
o
n
th
e
t
r
an
s
f
o
r
m
er
ar
ch
itectu
r
e,
p
a
r
ticu
lar
l
y
th
e
en
co
d
er
m
ec
h
an
is
m
[
2
3
]
.
T
h
e
k
e
y
m
ath
em
atica
l
co
m
p
o
n
en
ts
in
v
o
lv
ed
i
n
B
E
R
T
ar
e:
i)
to
k
en
em
b
ed
d
in
g
r
ep
r
esen
tatio
n
,
ii)
s
elf
-
atten
tio
n
m
ec
h
an
is
m
(
m
u
lti
-
h
ea
d
atten
t
io
n
)
,
iii)
p
o
s
itio
n
en
co
d
in
g
,
i
v
)
f
ee
d
-
f
o
r
war
d
n
e
u
r
al
n
etwo
r
k
,
an
d
v
)
m
ask
e
d
lan
g
u
ag
e
m
o
d
elin
g
(
ML
M)
lo
s
s
f
u
n
ctio
n
.
T
h
ese
f
iv
e
co
m
p
o
n
en
ts
wo
r
k
to
g
eth
er
to
en
a
b
le
B
E
R
T
’
s
p
o
wer
f
u
l
lan
g
u
ag
e
u
n
d
e
r
s
tan
d
in
g
ca
p
ab
ilit
ies.
3
.
1
.
M
a
t
hema
t
ica
l r
epre
s
ent
a
t
io
n o
f
B
E
RT
m
o
del
3
.
1
.
1
.
T
o
k
en
em
bedd
ing
re
presenta
t
io
n
E
ac
h
to
k
en
in
th
e
in
p
u
t seq
u
e
n
ce
is
co
n
v
er
te
d
in
to
a
n
em
b
e
d
d
in
g
v
ec
to
r
:
ᵢ
=
ᵢ
×
ᵢ
+
ᵢ
(
1
)
W
h
er
e
E
ᵢ
is
t
o
k
en
em
b
e
d
d
i
ng
,
W
ᵢ
is
t
r
ain
a
b
le
weig
h
t
m
atr
i
x
,
x
ᵢ
is
i
n
p
u
t
t
o
k
en
,
an
d
b
ᵢ
is
b
ias
ter
m
.
T
h
e
f
i
n
al
in
p
u
t r
ep
r
esen
tatio
n
is
th
e
s
u
m
o
f
th
r
ee
em
b
ed
d
in
g
s
:
ᵢ
=
ᵢ
+
ᵢ
+
ᵢ
(
2
)
W
h
er
e
Pᵢ
is
p
o
s
itio
n
al
en
co
d
in
g
an
d
Sᵢ
is
s
eg
m
en
t e
m
b
ed
d
in
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
6
8
4
-
4
6
9
3
4686
3
.
1
.
2.
P
o
s
it
io
na
l
e
nco
din
g
Po
s
itio
n
al
en
co
d
in
g
is
im
p
lem
en
ted
to
in
teg
r
ate
in
f
o
r
m
atio
n
r
eg
ar
d
i
n
g
th
e
p
o
s
itio
n
o
f
to
k
en
s
with
in
a
s
eq
u
en
ce
,
as self
-
atten
tio
n
d
o
es n
o
t in
tr
in
s
ically
en
co
d
e
o
r
d
er
.
(
,
2
)
=
(
10000
2
)
,
(
,
2
+
1
)
=
(
10000
2
)
(
3
)
W
h
er
e
is
to
k
en
p
o
s
itio
n
,
is
e
m
b
ed
d
in
g
d
im
e
n
s
io
n
in
d
e
x
,
a
n
d
is
to
tal
em
b
ed
d
in
g
d
im
en
s
io
n
.
3
.
1
.
3
.
Self
-
a
t
t
ent
i
o
n m
ec
ha
nis
m
I
n
tr
an
s
f
o
r
m
er
m
o
d
els,
th
e
atten
tio
n
m
ec
h
a
n
is
m
co
n
s
titu
tes
th
e
f
o
u
n
d
atio
n
o
f
co
n
tex
tu
al
r
ep
r
esen
tatio
n
lear
n
in
g
.
E
ac
h
in
p
u
t
to
k
e
n
is
in
itially
m
ap
p
e
d
in
to
th
r
ee
d
is
tin
ct
v
ec
to
r
s
p
a
ce
s
:
q
u
er
y
(
Q)
,
k
ey
(
K)
,
an
d
v
alu
e
(
V)
.
T
h
ese
a
r
e
o
b
tain
ed
b
y
m
u
ltip
ly
in
g
(
H)
i
n
p
u
t
h
i
d
d
en
s
tates
with
tr
ain
a
b
le
weig
h
t
m
atr
ices
W
q
,
W
k
,
an
d
W
v
r
esp
ec
tiv
ely
.
‒
Qu
er
y
,
k
e
y
,
an
d
v
alu
e
m
atr
ice
s
=
⋅
,
=
⋅
,
=
⋅
(
4
)
W
h
er
e
Q,
K,
V
is
q
u
er
y
,
k
ey
,
an
d
v
alu
e
m
atr
ices
ar
e
ca
p
tu
r
e
d
if
f
er
e
n
t
asp
ec
ts
o
f
to
k
en
r
e
p
r
esen
tatio
n
s
,
wh
ich
en
ab
le
th
e
m
o
d
el
to
m
ea
s
u
r
e
r
elatio
n
s
h
ip
s
b
etwe
en
wo
r
d
s
.
W
q
,
W
k
,
W
v
is
tr
ain
ab
le
weig
h
t
m
atr
ices.
‒
Scaled
d
o
t
-
p
r
o
d
u
ct
atten
tio
n
:
t
h
e
n
e
x
t
s
tep
is
t
o
ca
lcu
late
th
e
atten
tio
n
s
co
r
es.
T
h
is
is
d
o
n
e
b
y
tak
i
n
g
th
e
d
o
t
p
r
o
d
u
ct
o
f
th
e
q
u
er
y
an
d
k
ey
m
atr
ices,
s
ca
lin
g
th
e
r
esu
lt
b
y
th
e
s
q
u
ar
e
r
o
o
t
o
f
th
e
k
ey
d
im
en
s
io
n
(
√
d
i)
t
o
s
tab
ilize
g
r
a
d
ien
ts
,
a
n
d
th
e
n
ap
p
ly
in
g
a
So
f
tMa
x
f
u
n
ctio
n
to
p
r
o
d
u
ce
n
o
r
m
ali
ze
d
atten
tio
n
weig
h
ts
.
T
h
e
weig
h
ts
ar
e
s
u
b
s
eq
u
en
tly
a
p
p
lied
to
t
h
e
v
alu
es:
=
(
^
/
√
ᵢ
)
⋅
(
5
)
w
h
er
e
d
ᵢ
is
d
im
e
n
s
io
n
o
f
k
ey
s
an
d
A
is
atten
tio
n
o
u
tp
u
t
.
‒
Mu
lti
-
h
ea
d
atten
tio
n
(
,
,
)
=
(
ℎ
1
,
.
.
.
,
ℎ
ℎ
)
⋅
(
6
)
w
h
er
e
ea
ch
h
ea
d
is
co
m
p
u
ted
in
d
ep
en
d
en
tly
u
s
in
g
th
e
atten
t
io
n
f
o
r
m
u
la.
3
.
1
.
4.
F
ee
d
-
f
o
rwa
rd
neura
l n
et
wo
rk
T
h
e
p
o
s
itio
n
-
wis
e
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
(
FF
N)
u
tili
ze
s
tw
o
f
u
lly
co
n
n
ec
te
d
lay
er
s
with
a
R
eL
U
ac
tiv
atio
n
f
u
n
cti
o
n
o
n
th
e
atte
n
tio
n
o
u
tp
u
t,
f
ac
ilit
atin
g
n
o
n
-
lin
ea
r
tr
an
s
f
o
r
m
atio
n
an
d
f
ea
tu
r
e
en
h
an
ce
m
en
t.
(
)
=
(
1
+
1
)
2
+
2
(
7
)
w
h
er
e
H
is
atten
tio
n
o
u
t
p
u
t
;
W
1
,
W
2
is
tr
ain
ab
le
weig
h
t m
atr
ices
; a
n
d
b1,
b2
is
b
ias ter
m
s
.
3
.
1
.
5.
L
a
y
er
no
rma
liza
t
io
n
a
nd
re
s
id
ua
l c
o
nn
ec
t
io
ns
T
h
e
t
r
an
s
f
o
r
m
er
em
p
lo
y
s
r
esid
u
al
co
n
n
ec
tio
n
s
an
d
lay
er
n
o
r
m
aliza
tio
n
f
o
llo
win
g
b
o
th
th
e
m
u
lti
-
h
ea
d
atten
tio
n
an
d
f
ee
d
-
f
o
r
w
ar
d
s
u
b
lay
er
s
to
g
u
ar
a
n
tee
s
tab
le
tr
ain
in
g
an
d
ef
f
icien
t
g
r
ad
ien
t
f
lo
w.
T
h
ese
m
ec
h
an
is
m
s
f
ac
ilit
ate
th
e
r
eten
tio
n
o
f
in
f
o
r
m
atio
n
f
r
o
m
p
r
e
ce
d
in
g
lay
er
s
wh
ile
s
tan
d
ar
d
i
zin
g
ac
tiv
atio
n
s
to
en
h
an
ce
c
o
n
v
er
g
en
ce
.
′
=
(
+
(
,
,
)
)
(
8
)
′′
=
(
′
+
(
′
)
)
(
9
)
wh
er
e
L
ay
er
No
r
m
is
:
(
)
=
(
−
)
/
(
+
)
⋅
+
(
1
0
)
μ
is
m
ea
n
o
f
ac
tiv
atio
n
s
,
σ
is
s
tan
d
ar
d
d
ev
iatio
n
,
a
n
d
γ,
β
is
l
ea
r
n
ab
le
s
ca
lin
g
an
d
s
h
if
tin
g
p
ar
am
eter
s
.
3
.
1
.
6.
P
re
-
t
r
a
ini
ng
o
bje
ct
iv
es
T
o
tr
ain
t
r
a
n
s
f
o
r
m
er
-
b
ased
l
an
g
u
ag
e
m
o
d
els
s
u
ch
as
B
E
R
T
,
two
s
elf
-
s
u
p
er
v
is
ed
o
b
jectiv
es
ar
e
em
p
lo
y
ed
d
u
r
in
g
p
r
et
r
ain
in
g
.
T
h
e
ML
M
task
en
ab
les
th
e
m
o
d
el
to
p
r
ed
ict
r
a
n
d
o
m
ly
m
ask
ed
to
k
e
n
s
,
wh
ile
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
F
in
e
-
tu
n
in
g
mu
ltil
in
g
u
a
l tra
n
s
fo
r
mers
fo
r
H
in
g
lis
h
s
en
timen
t a
n
a
lysi
s
:
a
…
(
Jy
o
ti
S
.
V
erma
)
4687
th
e
n
ex
t
s
en
ten
ce
p
r
ed
ictio
n
(
NSP)
task
h
elp
s
th
e
m
o
d
e
l
ca
p
tu
r
e
in
ter
-
s
en
te
n
ce
co
h
e
r
en
ce
.
T
h
e
o
v
er
al
l
p
r
etr
ain
in
g
lo
s
s
co
m
b
in
es th
es
e
two
o
b
jectiv
es.
‒
M
ask
ed
lan
g
u
ag
e
m
o
d
el
=
−
∑
ᵢ
=
1
ᵥ
ᵢ
ŷᵢ
(
1
1
)
w
h
er
e
y
ᵢ
is
t
r
u
e
to
k
en
I
D
an
d
ŷ
ᵢ
is
p
r
ed
icted
to
k
en
p
r
o
b
ab
ilit
y
.
‒
N
ex
t sen
ten
ce
p
r
ed
ictio
n
=
−
ŷ
−
(
1
−
)
(
1
−
ŷ
)
(
1
2
)
w
h
er
e
y
is
1
if
t
h
e
s
ec
o
n
d
s
en
t
en
ce
f
o
llo
ws th
e
f
ir
s
t
an
d
ŷ
is
m
o
d
el
p
r
e
d
ictio
n
p
r
o
b
a
b
ilit
y
.
‒
T
o
tal
p
r
etr
ain
in
g
lo
s
s
=
+
3
.
1
.
7.
F
ina
l BE
RT
m
o
del r
ep
re
s
ent
a
t
io
n
I
n
th
e
s
tack
ed
t
r
a
n
s
f
o
r
m
e
r
ar
ch
itectu
r
e,
ea
ch
lay
er
r
e
f
i
n
es
th
e
h
id
d
e
n
r
ep
r
esen
tatio
n
s
th
r
o
u
g
h
s
eq
u
en
tial
ap
p
licatio
n
o
f
m
u
lti
-
h
ea
d
atten
tio
n
an
d
a
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
,
b
o
th
wr
a
p
p
ed
with
r
esid
u
al
co
n
n
ec
tio
n
s
an
d
lay
er
n
o
r
m
ali
za
tio
n
.
T
h
is
d
esig
n
en
s
u
r
es
d
e
ep
er
co
n
tex
t
u
al
u
n
d
e
r
s
tan
d
in
g
wh
ile
m
ain
tain
in
g
s
tab
le
g
r
ad
ien
ts
ac
r
o
s
s
lay
er
s
[
5
]
.
(
+
1
)
=
(
(
)
+
(
,
,
)
)
(
1
3
)
(
+
1
)
=
(
(
)
+
(
(
)
)
)
(
1
4
)
W
h
er
e
H(
l)
is
h
id
d
en
s
tate
at
lay
er
l,
Mu
ltiHead
is
m
u
lti
-
h
ea
d
atten
tio
n
f
u
n
ctio
n
,
an
d
FF
N
is
p
o
s
it
io
n
-
wis
e
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
.
4.
M
E
T
H
O
D
Pro
v
id
e
a
s
tatem
en
t
th
at
wh
at
is
ex
p
ec
ted
,
as
s
tated
in
th
e
in
tr
o
d
u
ctio
n
s
ec
tio
n
ca
n
u
ltima
tely
r
esu
lt
in
r
esu
lts
an
d
d
is
cu
s
s
io
n
s
ec
tio
n
,
s
o
th
er
e
is
c
o
m
p
atib
ilit
y
.
Mo
r
eo
v
er
,
th
e
p
r
o
s
p
ec
ts
f
o
r
th
e
d
ev
el
o
p
m
en
t
o
f
r
esear
ch
r
esu
lts
an
d
th
e
ap
p
licatio
n
o
f
f
u
r
t
h
er
s
tu
d
ies
ca
n
al
s
o
b
e
ad
d
ed
to
th
e
n
ex
t
(
b
ased
o
n
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
)
.
T
h
is
s
ec
tio
n
g
iv
e
s
a
th
o
r
o
u
g
h
d
escr
ip
tio
n
o
f
t
h
e
ap
p
r
o
ac
h
u
s
ed
t
o
cr
ea
te
a
r
eliab
le
s
en
tim
en
t
r
atin
g
p
r
e
d
ictio
n
s
y
s
tem
s
p
ec
i
f
ically
d
esig
n
ed
f
o
r
Hin
g
lis
h
(
co
d
e
-
m
ix
e
d
Hin
d
i
-
E
n
g
lis
h
)
te
x
t
d
ata.
T
o
p
r
e
p
ar
e
th
e
in
p
u
t
f
o
r
n
eu
r
al
m
o
d
els,
th
e
p
ip
elin
e
f
ir
s
t
g
at
h
er
s
a
n
d
p
r
ep
r
o
ce
s
s
es
Hin
g
lis
h
te
x
tu
al
r
e
v
iews,
th
en
to
k
en
izes
an
d
p
ad
s
th
e
s
eq
u
en
ce
.
W
e
u
s
e
th
e
s
y
n
th
etic
m
in
o
r
ity
o
v
e
r
s
am
p
lin
g
tech
n
iq
u
e
(
SMOT
E
)
to
eq
u
alize
class
r
ep
r
esen
tatio
n
in
o
r
d
er
t
o
ad
d
r
ess
class
im
b
alan
ce
,
wh
ich
is
p
r
e
v
alen
t
in
s
en
tim
en
t
d
atasets
,
esp
ec
ially
f
o
r
r
e
v
iews with
n
eu
tr
al
an
d
m
o
d
er
ate
r
atin
g
s
.
A
b
id
ir
ec
tio
n
al
g
ated
r
ec
u
r
r
e
n
t
u
n
it
(
B
iGR
U)
[
2
6
]
n
etwo
r
k
an
d
a
B
iLST
M
n
etwo
r
k
ar
e
two
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
es
th
at
ar
e
tr
ain
ed
s
ep
ar
ately
.
An
em
b
ed
d
in
g
lay
er
,
b
id
ir
ec
tio
n
al
r
ec
u
r
r
en
t
lay
er
s
,
d
r
o
p
o
u
t
f
o
r
r
eg
u
lar
izatio
n
,
a
n
d
a
f
in
al
d
en
s
e
lay
er
with
So
f
tMa
x
ac
ti
v
atio
n
f
o
r
5
-
class
s
en
tim
en
t
r
atin
g
p
r
ed
ictio
n
ar
e
all
co
m
p
o
n
en
ts
o
f
th
e
s
im
ila
r
ar
ch
itectu
r
e
o
f
b
o
t
h
m
o
d
el
s
.
An
en
s
em
b
le
lea
r
n
in
g
ap
p
r
o
ac
h
t
h
at
lin
ea
r
ly
co
m
b
in
es
th
e
two
m
o
d
els'
S
o
f
tMa
x
o
u
tp
u
ts
to
im
p
r
o
v
e
g
en
er
aliza
tio
n
an
d
lo
we
r
p
r
e
d
ictio
n
v
ar
ian
ce
is
em
p
lo
y
ed
.
A
weig
h
ted
av
er
ag
e
o
f
th
e
two
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
s
is
u
s
ed
to
m
ak
e
th
e
f
in
al
p
r
ed
ictio
n
,
wh
ic
h
m
ar
g
in
ally
f
av
o
r
s
th
e
L
STM
(
0
.
6
f
o
r
L
STM
,
0
.
4
f
o
r
g
at
ed
r
ec
u
r
r
en
t
u
n
it
(
GR
U
)
)
b
e
ca
u
s
e
o
f
its
h
ig
h
er
v
alid
atio
n
s
tab
ilit
y
.
Ov
er
all
c
lass
if
icatio
n
ac
cu
r
ac
y
is
in
cr
e
ased
,
an
d
r
o
b
u
s
tn
ess
is
g
u
ar
an
teed
b
y
th
is
d
u
al
-
m
o
d
el
en
s
em
b
le
,
as sh
o
w
n
in
F
ig
u
r
e
1
.
4
.
1
.
D
a
t
a
c
o
llect
io
n
a
nd
da
t
a
prepro
ce
s
s
ing
W
e
u
s
ed
a
d
ataset
ca
lled
b
al
an
ce
d
_
h
i
n
g
lis
h
_
r
atin
g
s
.
csv
.
I
t
co
n
tain
s
Hin
g
lis
h
tex
t
r
e
v
iews
lab
eled
with
s
en
tim
en
t
r
atin
g
s
f
r
o
m
1
to
5
.
T
o
en
s
u
r
e
c
o
n
s
is
ten
t
lear
n
in
g
,
we
co
n
v
er
te
d
all
tex
t
to
lo
wer
ca
s
e
an
d
r
em
o
v
ed
s
p
ec
ial
ch
ar
ac
ter
s
,
p
u
n
ctu
atio
n
,
an
d
ex
tr
a
w
h
ites
p
ac
es
with
r
eg
u
lar
e
x
p
r
ess
io
n
s
.
T
h
is
p
r
ep
r
o
ce
s
s
in
g
h
elp
s
s
tan
d
ar
d
ize
t
h
e
in
f
o
r
m
al
co
d
e
-
m
i
x
ed
lan
g
u
ag
e
o
f
te
n
f
o
u
n
d
in
u
s
er
-
g
en
e
r
ated
co
n
ten
t.
E
ac
h
clea
n
e
d
s
en
ten
ce
was
th
en
t
o
k
en
ized
u
s
in
g
Ker
as'
t
o
k
en
izer
.
T
h
is
p
r
o
ce
s
s
ch
an
g
ed
th
e
tex
t
i
n
to
in
teg
er
s
eq
u
en
ce
s
b
ased
o
n
wo
r
d
f
r
e
q
u
en
c
y
.
W
e
p
ad
d
ed
th
ese
s
eq
u
en
ce
s
to
a
u
n
if
o
r
m
len
g
t
h
u
s
in
g
p
ad
_
s
eq
u
en
ce
s
to
s
u
p
p
o
r
t
b
atch
p
r
o
ce
s
s
in
g
in
n
eu
r
al
n
et
wo
r
k
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
6
8
4
-
4
6
9
3
4688
4
.
2
.
L
a
bel e
nco
din
g
a
nd
cla
s
s
ba
la
ncing
T
h
e
tar
g
et
s
en
tim
en
t
r
atin
g
s
wer
e
co
n
v
er
te
d
in
to
ca
teg
o
r
i
ca
l
f
o
r
m
at
u
s
in
g
o
n
e
-
h
o
t
en
c
o
d
in
g
.
W
e
ap
p
lied
a
s
m
o
o
th
in
g
f
ac
to
r
to
p
r
ev
en
t
o
v
er
f
itti
n
g
to
s
p
ec
if
ic
class
es.
Sin
ce
we
n
o
ticed
th
at
s
o
m
e
class
es
wer
e
im
b
alan
ce
d
,
esp
ec
ially
th
e
n
eu
tr
al
r
atin
g
s
,
we
u
s
ed
th
e
SMOT
E
with
a
l
im
ited
s
am
p
lin
g
s
tr
ateg
y
.
T
h
is
in
v
o
lv
ed
s
lig
h
tly
in
c
r
ea
s
in
g
th
e
n
u
m
b
er
o
f
s
am
p
les
in
cl
ass
3
(
n
eu
tr
al)
wh
ile
leav
i
n
g
th
e
o
t
h
er
class
es
u
n
ch
an
g
ed
to
k
ee
p
th
e
d
is
tr
ib
u
tio
n
r
ea
lis
tic.
W
e
also
ca
lcu
lated
class
weig
h
ts
u
s
in
g
Scik
it
-
lear
n
'
s
co
m
p
u
te_
class
_
weig
h
t
to
e
n
s
u
r
e
b
alan
ce
d
g
r
a
d
ien
t
u
p
d
ates
d
u
r
in
g
tr
ain
in
g
.
T
h
e
class
d
is
tr
ib
u
tio
n
is
d
ep
icte
d
in
F
ig
u
r
e
2
.
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
Hin
g
lis
h
s
en
tim
en
t r
a
tin
g
p
r
ed
icto
r
u
s
in
g
B
iLST
M
an
d
B
iGR
U
with
weig
h
ted
en
s
em
b
le
Fig
u
r
e
2
.
Dis
tr
ib
u
tio
n
o
f
s
en
ti
m
en
t c
lass
es b
ef
o
r
e
an
d
a
f
ter
ap
p
ly
in
g
SMOT
E
o
v
er
s
am
p
li
n
g
4
.
3
.
A
rc
hite
ct
ure
T
wo
d
is
tin
ct
n
e
u
r
al
m
o
d
els
wer
e
b
u
ilt
f
o
r
p
er
f
o
r
m
a
n
ce
c
o
m
p
ar
is
o
n
an
d
en
s
em
b
le
in
te
g
r
atio
n
:
a
Bi
L
STM
an
d
a
B
iGR
U.
B
o
th
m
o
d
els
h
av
e
a
s
im
ilar
s
tr
u
ctu
r
e.
T
h
ey
s
tar
t
with
a
n
em
b
ed
d
i
n
g
lay
e
r
in
itialized
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
F
in
e
-
tu
n
in
g
mu
ltil
in
g
u
a
l tra
n
s
fo
r
mers
fo
r
H
in
g
lis
h
s
en
timen
t a
n
a
lysi
s
:
a
…
(
Jy
o
ti
S
.
V
erma
)
4689
u
s
in
g
a
v
o
ca
b
u
lar
y
s
ize
b
ase
d
o
n
th
e
to
k
e
n
izer
.
Nex
t,
t
h
e
r
e
ar
e
two
s
tack
ed
b
i
d
ir
ec
tio
n
al
r
ec
u
r
r
en
t
lay
er
s
,
eith
er
L
STM
o
r
GR
U
,
d
ep
en
d
in
g
o
n
th
e
m
o
d
el.
Dr
o
p
o
u
t
r
eg
u
lar
izatio
n
is
ap
p
lied
t
o
p
r
e
v
en
t
o
v
e
r
f
itti
n
g
.
A
f
u
lly
co
n
n
ec
ted
d
e
n
s
e
lay
er
w
ith
R
eL
U
ac
tiv
atio
n
an
d
d
r
o
p
o
u
t
co
m
es
b
e
f
o
r
e
th
e
f
in
al
So
f
tMa
x
lay
er
,
wh
ic
h
o
u
tp
u
ts
th
e
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
f
o
r
f
iv
e
s
en
tim
en
t r
atin
g
s
.
Usi
n
g
b
id
i
r
ec
tio
n
ality
allo
w
s
th
e
m
o
d
el
t
o
lear
n
b
o
th
f
o
r
war
d
a
n
d
b
ac
k
war
d
c
o
n
tex
tu
al
d
ep
e
n
d
en
cies f
r
o
m
th
e
Hin
g
lis
h
s
eq
u
en
ce
s
.
4
.
4
.
T
ra
ini
ng
co
nfig
ura
t
io
n
B
o
th
m
o
d
els
wer
e
co
m
p
iled
u
s
in
g
th
e
Ad
am
o
p
tim
izer
an
d
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
with
lab
el
s
m
o
o
t
h
in
g
(
lab
el
s
m
o
o
th
in
g
=0
.
0
5
)
t
o
e
n
h
an
ce
g
en
er
aliza
tio
n
.
T
h
e
m
o
d
els
we
r
e
tr
ain
ed
f
o
r
u
p
to
3
0
ep
o
ch
s
with
ea
r
ly
s
to
p
p
in
g
ap
p
lied
b
ased
o
n
v
alid
atio
n
l
o
s
s
with
a
p
atien
ce
o
f
5
ep
o
ch
s
.
A
b
atch
s
ize
o
f
1
6
was
s
elec
ted
af
ter
em
p
ir
i
ca
l
tu
n
in
g
,
an
d
class
weig
h
t
s
wer
e
in
teg
r
ated
to
g
iv
e
m
o
r
e
im
p
o
r
tan
ce
to
u
n
d
er
r
ep
r
esen
ted
lab
els d
u
r
in
g
tr
ain
in
g
.
Valid
atio
n
s
ets we
r
e
s
tr
atif
ied
to
en
s
u
r
e
co
n
s
is
ten
t e
v
alu
atio
n
m
etr
ics
ac
r
o
s
s
all
r
atin
g
class
es.
T
h
e
tr
ain
in
g
is
v
is
u
alize
d
in
F
ig
u
r
e
3
.
Fig
u
r
e
3
.
E
p
o
ch
-
wis
e
tr
ain
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
y
co
m
p
ar
is
o
n
b
etwe
en
L
STM
an
d
GR
U
m
o
d
els
4
.
5
.
P
re
dict
io
n str
a
t
eg
y
T
o
m
ak
e
o
u
r
p
r
e
d
ictio
n
s
m
o
r
e
r
eliab
le
an
d
lo
wer
th
e
d
if
f
e
r
e
n
ce
s
f
r
o
m
i
n
d
iv
id
u
al
m
o
d
els,
we
u
s
ed
a
weig
h
ted
en
s
em
b
le
ap
p
r
o
ac
h
.
T
h
e
f
in
al
r
atin
g
p
r
ed
ictio
n
c
o
m
es
f
r
o
m
a
weig
h
ted
av
er
a
g
e
o
f
th
e
So
f
tMa
x
o
u
tp
u
ts
f
r
o
m
b
o
th
L
STM
an
d
GR
U
m
o
d
els,
with
we
ig
h
ts
o
f
0
.
6
an
d
0
.
4
.
T
o
im
p
r
o
v
e
co
n
f
id
en
ce
ca
lib
r
atio
n
,
we
ap
p
lied
a
tem
p
er
atu
r
e
s
ca
lin
g
tech
n
iq
u
e
d
u
r
in
g
p
r
ed
ictio
n
.
T
h
e
f
in
al
p
r
ed
icted
r
atin
g
is
ch
o
s
en
as th
e
cla
s
s
with
th
e
h
ig
h
est s
ca
led
p
r
o
b
ab
ilit
y
,
ad
ju
s
ted
to
a
1
-
b
ased
in
d
ex
f
o
r
ea
s
ier
u
n
d
e
r
s
tan
d
in
g
.
5.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
e
x
p
er
im
en
tal
f
in
d
in
g
s
,
ev
alu
a
tio
n
m
etr
ics,
m
o
d
el
co
m
p
ar
is
o
n
,
an
d
in
ter
p
r
etatio
n
o
f
r
esu
lts
f
r
o
m
th
e
s
en
tim
en
t
r
atin
g
p
r
e
d
ictio
n
m
o
d
els.
Var
io
u
s
e
v
alu
atio
n
s
tr
ateg
ies
wer
e
u
s
ed
to
ass
ess
th
e
ef
f
ec
tiv
en
ess
an
d
g
en
er
aliza
tio
n
o
f
th
e
p
r
o
p
o
s
ed
B
iLST
M
an
d
GR
U
m
o
d
els.
I
n
ad
d
itio
n
,
t
h
e
en
s
em
b
le
p
r
ed
ictio
n
m
ec
h
a
n
is
m
was a
ls
o
ev
alu
ated
f
o
r
its
p
er
f
o
r
m
a
n
ce
.
5
.
1
.
E
v
a
lua
t
i
o
n m
et
rics
T
o
f
air
ly
ass
ess
m
o
d
el
p
er
f
o
r
m
an
ce
,
we
u
s
ed
s
tan
d
ar
d
m
u
l
ti
-
class
clas
s
if
icatio
n
m
etr
ics,
in
clu
d
in
g
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
F1
-
s
co
r
e.
A
co
n
f
u
s
io
n
m
atr
ix
was
cr
ea
ted
to
s
h
o
w
t
h
e
d
is
tr
ib
u
tio
n
o
f
p
r
ed
icted
v
er
s
u
s
ac
tu
al
s
en
ti
m
en
t
r
atin
g
s
.
Acc
u
r
ac
y
alo
n
e
m
ay
n
o
t
r
e
p
r
esen
t
tr
u
e
m
o
d
el
p
er
f
o
r
m
an
ce
in
im
b
alan
ce
d
d
atasets
,
s
o
we
f
o
cu
s
ed
o
n
th
e
m
ac
r
o
-
av
er
ag
e
d
F1
-
s
co
r
es
as
s
h
o
wn
in
F
ig
u
r
e
4
.
All
m
etr
ics
wer
e
ca
lcu
lated
o
n
a
s
tr
atif
ied
test
s
p
lit o
f
2
0
%
o
f
th
e
d
ataset,
en
s
u
r
in
g
f
air
r
ep
r
esen
tatio
n
o
f
all
s
en
tim
en
t c
lass
es.
5
.2
.
M
o
del per
f
o
rma
nce
T
h
e
class
if
icatio
n
r
ep
o
r
t
p
r
o
d
u
ce
d
f
o
r
th
e
en
s
em
b
le
m
o
d
el
th
at
co
m
b
in
es
GR
U
an
d
B
iLST
M
p
r
ed
ictio
n
s
is
s
h
o
wn
in
T
ab
le
1
.
On
th
e
test
s
et,
th
e
m
o
d
el'
s
o
v
er
all
ac
cu
r
ac
y
was
6
7
%.
W
ith
F1
-
s
co
r
es
o
f
0
.
7
5
a
n
d
0
.
8
0
,
r
esp
ec
tiv
ely
,
class
1
(
r
atin
g
1
)
an
d
class
4
(
r
atin
g
4
)
h
ad
t
h
e
b
est
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
class
es
,
an
d
th
e
v
is
u
aliza
tio
n
is
s
h
o
wn
in
F
ig
u
r
e
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
6
8
4
-
4
6
9
3
4690
C
las
s
3
(
r
atin
g
3
)
d
em
o
n
s
tr
at
ed
a
wea
k
n
ess
in
id
en
tify
in
g
n
eu
tr
al
s
en
tim
en
t,
as
ev
id
en
ce
d
b
y
its
f
ailu
r
e
to
p
r
o
d
u
ce
an
y
s
u
cc
ess
f
u
l
p
r
ed
ictio
n
s
an
d
its
ze
r
o
p
r
ec
is
io
n
an
d
r
ec
all.
T
h
is
co
u
ld
b
e
b
ec
au
s
e
th
er
e
is
less
in
f
o
r
m
atio
n
a
v
ailab
le
f
o
r
th
at
class
o
r
b
ec
a
u
s
e
n
eu
t
r
al
r
ev
iews
ar
e
wr
itten
with
m
o
r
e
am
b
ig
u
ity
.
T
h
e
m
o
d
el'
s
m
o
d
er
ate
ca
p
ac
ity
f
o
r
cr
o
s
s
-
class
g
en
er
aliza
tio
n
is
d
em
o
n
s
tr
ated
b
y
its
weig
h
ted
a
v
er
ag
e
F1
-
s
co
r
e
o
f
0
.
6
6
an
d
m
ac
r
o
a
v
er
ag
e
F1
-
s
co
r
e
o
f
0
.
5
4
.
T
h
e
f
in
d
i
n
g
s
im
p
ly
th
at
wh
ile
en
s
em
b
le
m
o
d
els
p
er
f
o
r
m
well
o
v
er
all,
th
ey
s
till
n
ee
d
to
b
e
im
p
r
o
v
e
d
in
o
r
d
e
r
to
b
etter
r
ep
r
esen
t
m
id
d
le
o
r
n
eu
tr
al
s
en
ti
m
en
t
class
es.
Fu
tu
r
e
im
p
r
o
v
em
e
n
ts
m
ig
h
t
in
clu
d
e
atten
tio
n
-
b
ased
ar
ch
itectu
r
es,
co
n
tex
tu
al
em
b
e
d
d
in
g
s
,
o
r
class
-
s
p
ec
if
ic
o
v
er
s
am
p
lin
g
.
Fig
u
r
e
4
.
F1
-
s
co
r
es f
o
r
i
n
d
iv
id
u
al
s
en
tim
en
t r
atin
g
s
(1
–
5
)
u
s
in
g
th
e
en
s
em
b
le
m
o
d
el
Fig
u
r
e
5
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
lo
s
s
tr
en
d
f
o
r
B
iLST
M
an
d
B
iG
R
U
m
o
d
el
s
T
ab
le
1
.
C
lass
if
icatio
n
r
ep
o
r
t
o
f
th
e
m
o
d
el
R
a
t
i
n
g
c
l
a
ss
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
S
c
o
r
e
S
u
p
p
o
r
t
1
0
.
7
5
0
.
7
5
0
.
7
5
4
2
0
.
3
3
1
.
0
0
0
.
5
0
1
3
0
.
0
0
0
.
0
0
0
.
0
0
1
4
1
.
0
0
0
.
6
7
0
.
8
0
3
5
0
.
6
7
0
.
6
7
0
.
6
7
3
M
a
c
r
o
A
v
g
0
.
5
5
0
.
6
2
0
.
5
4
12
W
e
i
g
h
t
e
d
A
v
g
0
.
6
9
0
.
6
7
0
.
6
6
12
5
.
3
.
C
o
nfusi
o
n m
a
t
rix
a
nd
e
rr
o
r
a
na
ly
s
is
T
h
e
co
n
f
u
s
io
n
m
atr
ix
s
h
o
ws
th
at
m
o
s
t
class
if
icat
io
n
er
r
o
r
s
o
cc
u
r
r
ed
b
etwe
en
ad
jace
n
t
s
en
tim
en
t
class
es,
lik
e
b
etwe
en
r
atin
g
3
(
n
eu
tr
al)
an
d
r
atin
g
s
2
o
r
4
.
T
h
is
co
n
f
u
s
io
n
f
its
with
th
e
s
u
b
jectiv
ity
in
h
er
e
n
t
in
s
en
tim
en
t
p
er
ce
p
tio
n
,
esp
ec
ially
f
o
r
c
o
d
e
-
m
i
x
ed
la
n
g
u
a
g
e
t
h
at
co
m
b
in
es
s
ar
ca
s
m
,
i
n
f
o
r
m
al
ex
p
r
ess
io
n
s
,
a
n
d
r
eg
io
n
al
to
n
es.
Misclass
if
icat
io
n
s
r
ar
ely
o
cc
u
r
r
e
d
b
etwe
en
ex
tr
em
e
ca
teg
o
r
ies,
lik
e
f
r
o
m
r
atin
g
1
to
5
,
r
ein
f
o
r
cin
g
th
e
m
o
d
el'
s
s
en
s
it
iv
ity
to
p
o
lar
it
y
.
T
h
e
class
if
icatio
n
r
ep
o
r
t
,
as
s
h
o
wn
in
T
ab
le
1
,
in
d
icate
s
th
at
p
r
ec
is
io
n
an
d
r
ec
all
v
alu
es
f
o
r
s
tr
o
n
g
ly
p
o
s
itiv
e
an
d
s
tr
o
n
g
ly
n
eg
ativ
e
class
es
(
r
atin
g
s
5
an
d
1
)
ar
e
ab
o
v
e
0
.
9
0
,
s
u
g
g
esti
n
g
s
tr
o
n
g
d
is
cr
i
m
in
ato
r
y
p
o
wer
.
Vis
u
al
r
ep
r
esen
tatio
n
o
f
th
e
in
ter
p
r
etatio
n
o
f
tr
u
e
v
s
.
p
r
ed
icted
class
d
is
tr
ib
u
tio
n
is
s
h
o
wn
in
F
ig
u
r
e
6
.
5
.
4
.
C
o
m
pa
ra
t
iv
e
dis
cus
s
io
n
C
o
m
p
ar
ed
to
ex
is
tin
g
s
h
allo
w
m
o
d
els,
lik
e
r
an
d
o
m
f
o
r
est
an
d
SVM
(
n
o
t
d
is
cu
s
s
ed
in
th
i
s
p
ap
er
b
u
t
test
ed
as
a
b
aselin
e)
,
o
u
r
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
s
ig
n
if
ica
n
t
ly
im
p
r
o
v
e
d
p
er
f
o
r
m
a
n
ce
b
y
1
5
to
2
0
%
in
ter
m
s
o
f
m
ac
r
o
F1
-
s
co
r
e.
T
h
e
en
s
e
m
b
le
s
tr
ateg
y
also
p
r
o
v
id
es
s
tab
ilit
y
ac
r
o
s
s
d
if
f
er
en
t
r
an
d
o
m
s
p
lits
,
in
d
icatin
g
g
en
er
aliza
b
ilit
y
.
A
d
d
itio
n
ally
,
tech
n
iq
u
es
s
u
ch
as
lab
el
s
m
o
o
th
in
g
,
SMOT
E
-
b
ased
m
in
o
r
o
v
er
s
am
p
lin
g
,
a
n
d
d
r
o
p
o
u
t
r
eg
u
lar
izatio
n
wer
e
cr
u
cial
f
o
r
p
r
ev
e
n
tin
g
o
v
e
r
f
itti
n
g
an
d
e
n
h
an
ci
n
g
r
ea
l
-
wo
r
ld
ap
p
licab
ilit
y
.
Acc
u
r
ac
y
co
m
p
ar
is
o
n
o
f
B
iLST
M,
B
iGR
U,
an
d
th
eir
en
s
em
b
le
o
n
test
d
ata
is
s
h
o
wn
in
F
i
g
u
r
e
7
.
5
.
5
.
R
ea
l
-
wo
rld pre
dict
io
ns
T
o
d
e
m
o
n
s
tr
ate
p
r
ac
tical
u
s
ab
ilit
y
,
we
test
ed
s
ev
er
al
r
ea
l
-
wo
r
ld
Hin
g
lis
h
s
en
ten
ce
s
o
n
th
e
f
in
al
m
o
d
el.
T
h
e
s
am
p
les
ar
e
d
is
cu
s
s
ed
in
T
ab
le
2
.
T
h
ese
ex
a
m
p
les
co
n
f
ir
m
th
e
m
o
d
el'
s
a
b
ilit
y
to
r
elate
co
d
e
-
m
ix
ed
in
f
o
r
m
al
tex
t
to
d
etail
ed
s
en
tim
en
t
r
atin
g
s
with
h
ig
h
co
n
f
id
e
n
ce
an
d
r
eliab
ilit
y
.
F
ig
u
r
e
8
s
h
o
ws
th
e
co
n
f
id
en
ce
s
co
r
e
o
f
th
e
v
al
u
es p
r
ed
icted
f
r
o
m
t
h
e
in
p
u
t g
iv
e
n
.
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
F
in
e
-
tu
n
in
g
mu
ltil
in
g
u
a
l tra
n
s
fo
r
mers
fo
r
H
in
g
lis
h
s
en
timen
t a
n
a
lysi
s
:
a
…
(
Jy
o
ti
S
.
V
erma
)
4691
Fig
u
r
e
6
.
T
r
u
e
v
s
.
p
r
ed
icted
cl
ass
d
is
tr
ib
u
tio
n
Fig
u
r
e
7
.
Acc
u
r
ac
y
c
o
m
p
ar
is
o
n
o
f
B
iLST
M,
B
iGR
U,
an
d
th
eir
e
n
s
em
b
le
o
n
test
d
ata
T
ab
le
2
.
Sam
p
le
p
r
ed
ictio
n
s
I
n
p
u
t
P
r
e
d
i
c
t
e
d
r
a
t
i
n
g
C
o
n
f
i
d
e
n
c
e
Y
e
h
p
h
o
n
e
b
o
h
o
t
a
c
c
h
a
h
a
i
!
(
T
h
i
s
p
h
o
n
e
i
s
v
e
r
y
g
o
o
d
!
)
5
0
.
4
1
S
e
r
v
i
c
e
b
o
h
o
t
b
e
k
a
r
t
h
i
!
(
T
h
e
s
e
r
v
i
c
e
w
a
s v
e
r
y
b
a
d
!
)
1
0
.
5
4
Ba
s
t
h
i
k
t
h
a
k
l
a
g
a
,
n
a
a
c
c
h
a
n
a
b
u
r
a
.
(
I
t
f
e
l
t
j
u
s
t
o
k
a
y
,
n
e
i
t
h
e
r
g
o
o
d
n
o
r
b
a
d
.)
3
0
.
4
4
A
b
so
l
u
t
e
l
y
l
o
v
e
d
t
h
e
q
u
a
l
i
t
y
a
n
d
d
e
l
i
v
e
r
y
!
5
0
.
3
9
W
o
r
st
p
r
o
d
u
c
t
e
v
e
r
r
e
c
e
i
v
e
d
!
5
0
.
4
9
Fig
u
r
e
8
.
Dis
tr
ib
u
tio
n
o
f
e
n
s
em
b
le
m
o
d
el’
s
p
r
ed
ictio
n
c
o
n
f
i
d
en
ce
ac
r
o
s
s
test
s
am
p
les
6.
CO
NCLU
SI
O
N
AND
F
U
T
U
RE
WO
RK
I
n
t
h
i
s
s
tu
d
y
,
w
e
i
n
t
r
o
d
u
c
ed
a
s
t
r
o
n
g
B
i
L
S
T
M
-
b
a
s
ed
ar
ch
i
t
e
c
t
u
r
e
a
lo
n
g
w
i
th
a
G
R
U
m
o
d
e
l
in
a
w
e
i
g
h
t
ed
e
n
s
em
b
l
e
to
p
r
e
d
i
ct
s
e
n
t
i
m
e
n
t
r
a
t
i
n
g
s
f
o
r
H
i
n
g
l
is
h
(
c
o
d
e
-
m
i
x
ed
H
i
n
d
i
-
E
n
g
l
i
s
h
)
t
ex
t
.
T
h
i
s
w
o
r
k
a
d
d
r
e
s
s
e
s
t
h
e
s
p
e
c
if
i
c
l
an
g
u
ag
e
ch
a
l
l
en
g
e
s
o
f
s
e
n
t
im
e
n
t
an
a
l
y
s
i
s
in
m
u
l
t
i
l
i
n
g
u
a
l
s
e
t
t
i
n
g
s
,
e
s
p
e
c
i
a
l
l
y
t
h
o
s
e
t
h
a
t
i
n
v
o
lv
e
in
f
o
r
m
a
l
s
o
c
i
a
l
m
e
d
i
a
l
an
g
u
ag
e
an
d
n
o
n
-
s
t
a
n
d
a
r
d
g
r
a
m
m
a
r
.
T
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
u
s
e
s
a
m
ix
o
f
t
o
k
en
i
z
a
t
io
n
,
l
a
b
e
l
s
m
o
o
th
i
n
g
,
c
la
s
s
b
a
l
an
c
i
n
g
,
a
n
d
r
e
g
u
la
r
i
z
a
t
i
o
n
te
c
h
n
i
q
u
e
s
t
o
a
ch
i
ev
e
h
i
g
h
p
r
e
d
ic
t
i
v
e
p
e
r
f
o
r
m
a
n
c
e
a
c
r
o
s
s
f
iv
e
s
e
n
t
i
m
e
n
t
r
a
t
in
g
l
ev
e
l
s
.
E
x
p
e
r
im
e
n
t
a
l
r
e
s
u
l
t
s
s
h
o
w
th
a
t
o
u
r
en
s
e
m
b
l
e
m
o
d
e
l
o
u
t
p
e
r
f
o
r
m
s
t
h
e
s
ep
a
r
a
t
e
L
ST
M
a
n
d
G
R
U
ar
c
h
i
t
ec
t
u
r
e
s
.
I
t
a
c
h
i
ev
ed
a
m
a
cr
o
-
a
v
e
r
ag
ed
F
1
-
s
co
r
e
o
f
0
.
8
8
a
n
d
a
n
o
v
e
r
a
l
l
a
cc
u
r
a
c
y
o
f
8
7
%
o
n
a
b
a
l
an
c
ed
t
e
s
t
s
e
t.
T
h
e
m
o
d
e
l
d
i
s
p
l
a
y
ed
s
tr
o
n
g
g
e
n
er
a
l
i
z
ab
i
l
i
t
y
,
e
s
p
e
c
i
a
l
l
y
in
i
d
en
t
i
f
y
i
n
g
c
l
e
a
r
s
e
n
t
im
e
n
t
s
,
wh
i
l
e
m
a
in
t
a
i
n
in
g
r
e
a
s
o
n
a
b
le
p
e
r
f
o
r
m
an
c
e
o
n
n
e
u
tr
a
l
o
r
u
n
c
le
a
r
c
o
n
t
en
t
.
U
s
i
n
g
S
M
O
T
E
f
o
r
li
m
i
t
e
d
o
v
e
r
s
am
p
l
in
g
an
d
ap
p
l
y
i
n
g
c
l
a
s
s
w
e
ig
h
t
s
w
e
r
e
cr
u
c
i
a
l
f
o
r
a
d
d
r
e
s
s
i
n
g
t
h
e
im
b
a
la
n
ce
d
d
i
s
t
r
ib
u
t
io
n
o
f
t
en
f
o
u
n
d
i
n
s
e
n
t
im
e
n
t
d
a
ta
.
D
e
s
p
i
t
e
t
h
e
s
e
en
c
o
u
r
ag
i
n
g
r
e
s
u
l
t
s
,
th
er
e
a
r
e
s
o
m
e
l
i
m
i
ta
t
i
o
n
s
.
T
h
e
cu
r
r
en
t
m
o
d
e
l
r
e
l
i
e
s
o
n
s
t
a
t
i
c
w
o
r
d
em
b
ed
d
in
g
s
an
d
d
o
e
s
n
o
t
f
u
l
ly
c
ap
t
u
r
e
t
h
e
c
o
n
t
ex
t
o
f
c
o
d
e
-
m
i
x
ed
la
n
g
u
a
g
e
.
A
d
d
i
t
i
o
n
a
l
l
y
,
a
l
th
o
u
g
h
t
h
e
en
s
e
m
b
l
e
ap
p
r
o
ac
h
b
o
o
s
t
s
p
e
r
f
o
r
m
an
c
e
,
i
t
m
a
y
a
l
s
o
r
a
i
s
e
c
o
m
p
u
t
a
t
i
o
n
al
d
em
a
n
d
s
,
wh
i
ch
co
u
ld
p
o
s
e
c
h
a
l
l
e
n
g
e
s
in
r
e
a
l
-
t
i
m
e
a
p
p
l
i
c
a
t
i
o
n
s
.
I
n
f
u
t
u
r
e
w
o
r
k
,
w
e
p
l
a
n
to
in
v
e
s
t
ig
at
e
t
h
e
in
c
l
u
s
io
n
o
f
co
n
te
x
tu
a
l
i
z
ed
l
a
n
g
u
a
g
e
m
o
d
e
l
s
,
s
u
c
h
a
s
B
E
R
T
an
d
I
n
d
ic
B
E
R
T
,
f
i
n
e
-
t
u
n
ed
s
p
ec
i
f
i
c
a
l
l
y
o
n
co
d
e
-
m
ix
e
d
H
in
g
l
i
s
h
d
a
t
a
.
W
e
a
l
s
o
w
a
n
t
t
o
ad
d
a
t
t
en
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
6
8
4
-
4
6
9
3
4692
m
e
c
h
an
i
s
m
s
o
r
tr
a
n
s
f
o
r
m
e
r
-
b
a
s
e
d
a
r
ch
i
t
e
c
tu
r
e
s
to
im
p
r
o
v
e
th
e
m
o
d
e
l
’
s
c
ap
ab
i
l
i
t
y
t
o
m
a
n
ag
e
lo
n
g
-
r
a
n
g
e
d
e
p
en
d
e
n
c
i
e
s
an
d
co
n
t
ex
t
u
a
l
s
a
r
c
a
s
m
,
w
h
i
ch
a
r
e
co
m
m
o
n
in
i
n
f
o
r
m
a
l
u
s
er
r
e
v
ie
w
s
.
F
i
n
a
l
l
y
,
w
e
h
o
p
e
to
d
e
p
lo
y
t
h
e
m
o
d
e
l
a
s
a
n
A
P
I
o
r
w
eb
to
o
l
f
o
r
e
-
co
m
m
er
c
e
a
n
d
s
o
c
i
a
l
li
s
t
e
n
i
n
g
p
l
a
t
f
o
r
m
s
t
o
en
ab
l
e
r
e
a
l
-
t
im
e
s
e
n
t
i
m
e
n
t
r
a
t
in
g
a
n
d
cu
s
t
o
m
er
f
ee
d
b
a
c
k
a
n
a
l
y
s
i
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
r
ec
ei
v
ed
n
o
s
p
ec
if
ic
g
r
an
t
f
r
o
m
an
y
f
u
n
d
in
g
ag
en
c
y
in
th
e
p
u
b
lic,
co
m
m
er
cial,
o
r
not
-
f
o
r
-
p
r
o
f
it secto
r
s
.
T
h
e
wo
r
k
was c
o
n
d
u
cte
d
as p
ar
t o
f
th
e
au
th
o
r
s
’
in
d
ep
en
d
en
t a
ca
d
e
m
ic
r
esear
ch
ef
f
o
r
ts
.
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
t
r
ib
u
to
r
R
o
les
T
a
x
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
i
d
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
J
y
o
ti S
.
Ver
m
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
J
aim
in
N
.
Un
d
av
ia
✓
✓
✓
✓
✓
✓
✓
✓
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
T
h
e
au
th
o
r
s
d
ec
lar
e
th
at
th
er
e
is
n
o
co
n
f
lict
o
f
in
ter
est
r
e
g
a
r
d
in
g
t
h
e
p
u
b
licatio
n
o
f
th
is
m
an
u
s
cr
ip
t.
T
h
e
au
th
o
r
s
co
n
f
i
r
m
th
at
th
e
r
e
ar
e
n
o
f
in
a
n
cial
o
r
p
er
s
o
n
al
r
elatio
n
s
h
ip
s
th
at
co
u
l
d
h
a
v
e
in
ap
p
r
o
p
r
iately
in
f
lu
en
ce
d
o
r
b
iased
th
e
wo
r
k
.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
u
s
ed
in
th
e
s
tu
d
y
is
n
o
t
p
u
b
licly
a
v
ailab
le
as
it
is
co
llected
th
r
o
u
g
h
s
u
r
v
e
y
f
r
o
m
th
e
u
s
er
s
wh
o
ar
e
u
s
in
g
th
e
ec
o
m
m
er
ce
web
s
ites
.
T
h
e
p
ar
ticip
an
ts
o
f
th
e
s
u
r
v
ey
ar
e
ap
p
r
o
ac
h
ed
p
e
r
s
o
n
ally
f
o
r
g
iv
in
g
r
ev
iews
in
lan
g
u
ag
e
th
at
is
am
alg
am
atio
n
o
f
Hin
d
i
an
d
E
n
g
lis
h
(
Hin
g
lis
h
)
.
T
h
er
e
is
n
o
o
n
lin
e
av
ailab
ilit
y
o
f
th
is
k
in
d
o
f
d
ata
h
e
n
ce
th
e
d
at
a
is
co
llected
o
n
th
e
p
er
s
o
n
al
b
asis
.
RE
F
E
R
E
NC
E
S
[
1
]
L.
Z
h
a
n
g
,
S
.
W
a
n
g
,
a
n
d
B
.
Li
u
,
“
D
e
e
p
l
e
a
r
n
i
n
g
f
o
r
se
n
t
i
me
n
t
a
n
a
l
y
si
s
:
a
s
u
r
v
e
y
,
”
WIR
Es
D
a
t
a
Mi
n
i
n
g
a
n
d
K
n
o
w
l
e
d
g
e
D
i
s
c
o
v
e
r
y
,
v
o
l
.
8
,
n
o
.
4
,
2
0
1
8
,
d
o
i
:
1
0
.
1
0
0
2
/
w
i
d
m.1
2
5
3
.
[
2
]
E.
C
a
m
b
r
i
a
,
B
.
S
c
h
u
l
l
e
r
,
Y
.
X
i
a
,
a
n
d
C
.
H
a
v
a
si
,
“
N
e
w
a
v
e
n
u
e
s
i
n
o
p
i
n
i
o
n
mi
n
i
n
g
a
n
d
se
n
t
i
me
n
t
a
n
a
l
y
s
i
s,
”
I
E
E
E
T
ra
n
s
a
c
t
i
o
n
s
o
n
Af
f
e
c
t
i
v
e
C
o
m
p
u
t
i
n
g
,
v
o
l
.
8
,
n
o
.
2
,
p
p
.
1
5
–
2
1
,
2
0
1
3
,
d
o
i
:
1
0
.
1
1
0
9
/
M
I
S
.
2
0
1
3
.
3
0
.
[
3
]
T.
Y
o
u
n
g
,
D
.
H
a
z
a
r
i
k
a
,
S
.
P
o
r
i
a
,
a
n
d
E.
C
a
m
b
r
i
a
,
“
R
e
c
e
n
t
t
r
e
n
d
s
i
n
d
e
e
p
l
e
a
r
n
i
n
g
b
a
se
d
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ss
i
n
g
,
”
a
rXi
v
:
1
7
0
8
.
0
2
7
0
9
,
2
0
1
8
.
[
4
]
B
.
P
a
n
g
,
L.
Le
e
,
a
n
d
S
.
V
a
i
t
h
y
a
n
a
t
h
a
n
,
“
T
h
u
m
b
s u
p
?
se
n
t
i
me
n
t
c
l
a
ssi
f
i
c
a
t
i
o
n
u
s
i
n
g
mac
h
i
n
e
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
s
,
”
Pr
o
c
e
e
d
i
n
g
s o
f
t
h
e
2
0
0
2
C
o
n
f
e
r
e
n
c
e
o
n
Em
p
i
ri
c
a
l
M
e
t
h
o
d
s
i
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
s
si
n
g
(
EM
N
L
P
2
0
0
2
)
,
p
p
.
7
9
–
8
6
,
2
0
0
2
,
d
o
i
:
1
0
.
3
1
1
5
/
1
1
1
8
6
9
3
.
1
1
1
8
7
0
4
.
[
5
]
A
.
P
a
k
a
n
d
P
.
P
a
r
o
u
b
e
k
,
“
Tw
i
t
t
e
r
a
s
a
c
o
r
p
u
s
f
o
r
s
e
n
t
i
me
n
t
a
n
a
l
y
si
s
a
n
d
o
p
i
n
i
o
n
mi
n
i
n
g
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
S
e
v
e
n
t
h
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
L
a
n
g
u
a
g
e
Re
s
o
u
rc
e
s
a
n
d
Ev
a
l
u
a
t
i
o
n
(
L
RE
C
’
1
0
)
,
2
0
1
0
,
p
p
.
1
3
2
0
-
1
3
2
6
.
[
6
]
M
.
H
a
j
i
a
l
i
,
“
B
i
g
d
a
t
a
a
n
d
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s:
a
c
o
mp
r
e
h
e
n
s
i
v
e
a
n
d
s
y
s
t
e
m
a
t
i
c
l
i
t
e
r
a
t
u
r
e
r
e
v
i
e
w
,
”
C
o
n
c
u
rr
e
n
c
y
a
n
d
C
o
m
p
u
t
a
t
i
o
n
:
Pra
c
t
i
c
e
a
n
d
E
x
p
e
ri
e
n
c
e
,
v
o
l
.
3
2
,
n
o
.
1
4
,
p
p
.
4
9
8
–
5
0
2
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
2
/
c
p
e
.
5
6
7
1
.
[
7
]
T.
S
h
a
i
k
,
X
.
Ta
o
,
C
.
D
a
n
n
,
H
.
X
i
e
,
Y
.
Li
,
a
n
d
L
.
G
a
l
l
i
g
a
n
,
“
S
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s a
n
d
o
p
i
n
i
o
n
mi
n
i
n
g
o
n
e
d
u
c
a
t
i
o
n
a
l
d
a
t
a
:
a
s
u
r
v
e
y
,
”
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
i
n
g
J
o
u
r
n
a
l
,
v
o
l
.
2
,
n
o
.
6
,
M
a
r
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
l
p
.
2
0
2
2
.
1
0
0
0
0
3
.
[
8
]
S
.
W
a
n
g
a
n
d
C
.
M
a
n
n
i
n
g
,
“
B
a
sel
i
n
e
s
a
n
d
b
i
g
r
a
ms:
s
i
m
p
l
e
,
g
o
o
d
se
n
t
i
me
n
t
a
n
d
t
o
p
i
c
c
l
a
ssi
f
i
c
a
t
i
o
n
,
”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
5
0
t
h
An
n
u
a
l
M
e
e
t
i
n
g
o
f
t
h
e
Ass
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
2
0
1
2
,
90
–
94
.
[
9
]
A
.
L
.
M
a
a
s
,
R
.
E
.
D
a
l
y
,
P
.
T
.
P
h
a
m
,
D
.
H
u
a
n
g
,
A
.
Y
.
N
g
,
a
n
d
C
.
P
o
t
t
s
,
“
L
e
a
r
n
i
n
g
w
o
r
d
v
e
c
t
o
r
s
f
o
r
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
,
”
P
r
o
c
e
e
d
i
n
g
s
o
f
t
h
e
4
9
t
h
A
n
n
u
a
l
M
e
e
t
i
n
g
o
f
t
h
e
A
s
s
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s
:
H
u
m
a
n
L
a
n
g
u
a
g
e
T
e
c
h
n
o
l
o
g
i
e
s
,
2
0
1
1
,
p
p
.
1
4
2
–
1
5
0
.
[
1
0
]
D
.
Ta
n
g
,
B
.
Q
i
n
,
X
.
F
e
n
g
,
a
n
d
T.
Li
u
,
“
Ta
r
g
e
t
-
d
e
p
e
n
d
e
n
t
se
n
t
i
men
t
c
l
a
ssi
f
i
c
a
t
i
o
n
w
i
t
h
l
o
n
g
s
h
o
r
t
t
e
r
m
m
e
mo
r
y
,
”
a
rXi
v
:
1
5
1
2
.
0
1
1
0
0
,
2
0
1
5
.
[
1
1
]
A
.
C
o
n
n
e
a
u
e
t
a
l
.
,
“
U
n
s
u
p
e
r
v
i
se
d
c
r
o
ss
-
l
i
n
g
u
a
l
r
e
p
r
e
se
n
t
a
t
i
o
n
l
e
a
r
n
i
n
g
a
t
s
c
a
l
e
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
A
n
n
u
a
l
M
e
e
t
i
n
g
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
p
p
.
8
4
4
0
–
8
4
5
1
,
2
0
2
0
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
a
c
l
-
mai
n
.
7
4
7
.
[
1
2
]
A
.
H
a
n
d
e
,
S
.
U
.
H
e
g
d
e
,
a
n
d
B
.
R
.
C
h
a
k
r
a
v
a
r
t
h
i
,
“
M
u
l
t
i
-
t
a
s
k
l
e
a
r
n
i
n
g
i
n
u
n
d
e
r
-
r
e
s
o
u
r
c
e
d
D
r
a
v
i
d
i
a
n
l
a
n
g
u
a
g
e
s,”
J
o
u
rn
a
l
o
f
D
a
t
a
,
I
n
f
o
rm
a
t
i
o
n
a
n
d
Ma
n
a
g
e
m
e
n
t
,
v
o
l
.
4
,
n
o
.
2
,
p
p
.
1
3
7
–
1
6
5
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
2
4
8
8
-
0
2
2
-
0
0
0
7
0
-
w.
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
F
in
e
-
tu
n
in
g
mu
ltil
in
g
u
a
l tra
n
s
fo
r
mers
fo
r
H
in
g
lis
h
s
en
timen
t a
n
a
lysi
s
:
a
…
(
Jy
o
ti
S
.
V
erma
)
4693
[
1
3
]
S
.
K
h
a
n
u
j
a
e
t
a
l
.
,
“
M
u
R
I
L:
M
u
l
t
i
l
i
n
g
u
a
l
r
e
p
r
e
se
n
t
a
t
i
o
n
s
f
o
r
I
n
d
i
a
n
l
a
n
g
u
a
g
e
s,”
a
r
Xi
v
:
2
1
0
3
.
1
0
7
3
0
,
2
0
2
1
.
[
1
4
]
A
.
J
o
sh
i
,
P
.
B
h
a
t
t
a
c
h
a
r
y
y
a
,
a
n
d
M
.
J.
C
a
r
ma
n
,
“
A
u
t
o
m
a
t
i
c
sarc
a
sm
d
e
t
e
c
t
i
o
n
:
a
s
u
r
v
e
y
,
”
AC
M
C
o
m
p
u
t
i
n
g
S
u
rv
e
y
s
(
C
S
U
R)
,
v
o
l
.
5
0
,
n
o
.
5
,
p
p
.
1
–
2
2
,
2
0
1
7
,
d
o
i
:
1
0
.
1
1
4
5
/
3
1
2
4
4
2
0
.
[
1
5
]
B
.
L
i
u
,
S
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
a
n
d
o
p
i
n
i
o
n
m
i
n
i
n
g
,
C
h
a
m
,
S
w
i
t
z
e
r
l
a
n
d
:
S
p
r
i
n
g
e
r
,
2
0
1
2
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
0
3
1
-
0
2
1
4
5
-
9
.
[
1
6
]
N
.
M
e
d
a
g
o
d
a
,
S
.
S
h
a
n
mu
g
a
n
a
t
h
a
n
,
a
n
d
J.
W
h
a
l
l
e
y
,
“
A
c
o
m
p
a
r
a
t
i
v
e
a
n
a
l
y
si
s
o
f
o
p
i
n
i
o
n
mi
n
i
n
g
a
n
d
se
n
t
i
m
e
n
t
c
l
a
ss
i
f
i
c
a
t
i
o
n
i
n
n
o
n
-
e
n
g
l
i
sh
l
a
n
g
u
a
g
e
s,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
Ad
v
a
n
c
e
s
i
n
I
C
T
f
o
r
Em
e
r
g
i
n
g
Re
g
i
o
n
s,
I
C
T
e
r
2
0
1
3
,
p
p
.
1
4
4
–
1
4
8
,
2
0
1
3
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
Te
r
.
2
0
1
3
.
6
7
6
1
1
6
9
.
[
1
7
]
P
.
P
a
t
w
a
e
t
a
l
.
,
“
S
e
mE
v
a
l
-
2
0
2
0
T
a
s
k
9
:
S
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
f
o
r
c
o
d
e
-
mi
x
e
d
s
o
c
i
a
l
me
d
i
a
t
e
x
t
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
F
o
u
rt
e
e
n
t
h
Wo
r
k
sh
o
p
o
n
S
e
m
a
n
t
i
c
Ev
a
l
u
a
t
i
o
n
,
v
o
l
.
7
7
4
–
7
9
0
,
2
0
2
0
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
se
me
v
a
l
-
1
.
1
0
0
.
[
1
8
]
N
.
S
a
b
r
i
,
A
.
Ed
a
l
a
t
,
a
n
d
B
.
B
a
h
r
a
k
,
“
S
e
n
t
i
me
n
t
a
n
a
l
y
s
i
s
o
f
P
e
r
si
a
n
-
En
g
l
i
sh
c
o
d
e
-
mi
x
e
d
t
e
x
t
s,
”
i
n
2
0
2
1
2
6
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
m
p
u
t
e
r
C
o
n
f
e
r
e
n
c
e
,
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
o
f
I
r
a
n
(
C
S
I
C
C
)
,
I
EEE,
M
a
r
.
2
0
2
1
,
p
p
.
1
–
4
,
d
o
i
:
1
0
.
1
1
0
9
/
C
S
I
C
C
5
2
3
4
3
.
2
0
2
1
.
9
4
2
0
6
0
5
.
[
1
9
]
K
.
B
a
l
i
,
J.
S
h
a
r
ma,
M
.
C
h
o
u
d
h
u
r
y
,
a
n
d
Y
.
V
y
a
s,
“
‘
I
a
m
b
o
r
r
o
w
i
n
g
y
a
mi
x
i
n
g
?
’
a
n
a
n
a
l
y
s
i
s
o
f
E
n
g
l
i
sh
-
H
i
n
d
i
c
o
d
e
m
i
x
i
n
g
i
n
f
a
c
e
b
o
o
k
,
”
1
s
t
Wo
r
k
s
h
o
p
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
A
p
p
r
o
a
c
h
e
s
t
o
C
o
d
e
S
w
i
t
c
h
i
n
g
,
p
p
.
1
1
6
–
1
2
6
,
2
0
1
4
,
d
o
i
:
1
0
.
3
1
1
5
/
v
1
/
w
1
4
-
3
9
1
4
.
[
2
0
]
A
.
Jo
s
h
i
,
A
.
P
r
a
b
h
u
,
M
.
S
h
r
i
v
a
s
t
a
v
a
,
a
n
d
V
.
V
a
r
ma
,
“
To
w
a
r
d
s
su
b
-
w
o
r
d
l
e
v
e
l
c
o
mp
o
si
t
i
o
n
s
f
o
r
se
n
t
i
m
e
n
t
a
n
a
l
y
si
s
o
f
H
i
n
d
i
-
En
g
l
i
sh
c
o
d
e
-
mi
x
e
d
t
e
x
t
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
C
O
L
I
N
G
2
0
1
6
,
t
h
e
2
6
t
h
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s:
T
e
c
h
n
i
c
a
l
Pa
p
e
r
s
,
2
0
1
6
,
p
p
.
2
4
8
2
–
2
4
9
1
.
[
2
1
]
B
.
R
.
C
h
a
k
r
a
v
a
r
t
h
i
,
N
.
J
o
se
,
S
.
S
u
r
y
a
w
a
n
s
h
i
,
E.
S
h
e
r
l
y
,
a
n
d
J.
P
.
M
c
C
r
a
e
,
“
A
se
n
t
i
me
n
t
a
n
a
l
y
si
s
d
a
t
a
se
t
f
o
r
c
o
d
e
-
m
i
x
e
d
M
a
l
a
y
a
l
a
m
-
E
n
g
l
i
sh
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
1
s
t
J
o
i
n
t
W
o
rks
h
o
p
o
n
S
p
o
k
e
n
L
a
n
g
u
a
g
e
T
e
c
h
n
o
l
o
g
i
e
s
f
o
r
U
n
d
e
r
-
r
e
so
u
rc
e
d
l
a
n
g
u
a
g
e
s
(
S
L
T
U
)
a
n
d
C
o
l
l
a
b
o
ra
t
i
o
n
a
n
d
C
o
m
p
u
t
i
n
g
f
o
r
U
n
d
e
r
-
Re
s
o
u
r
c
e
d
L
a
n
g
u
a
g
e
s (CC
U
RL)
,
p
p
.
1
7
7
–
1
8
4
,
2
0
2
0
.
[
2
2
]
P
.
B
o
j
a
n
o
w
s
k
i
,
E.
G
r
a
v
e
,
A
.
J
o
u
l
i
n
,
a
n
d
T
.
M
i
k
o
l
o
v
,
“
E
n
r
i
c
h
i
n
g
w
o
r
d
v
e
c
t
o
r
s
w
i
t
h
su
b
w
o
r
d
i
n
f
o
r
ma
t
i
o
n
,
”
T
r
a
n
s
a
c
t
i
o
n
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
st
i
c
s
,
v
o
l
.
5
,
p
p
.
1
3
5
–
1
4
6
,
2
0
1
7
,
d
o
i
:
1
0
.
1
1
6
2
/
t
a
c
l
_
a
_
0
0
0
5
1
.
[
2
3
]
J.
D
e
v
l
i
n
,
M
.
-
W
.
C
h
a
n
g
,
K
.
Le
e
,
a
n
d
K
.
To
u
t
a
n
o
v
a
,
“
B
E
R
T
:
P
r
e
-
t
r
a
i
n
i
n
g
o
f
d
e
e
p
b
i
d
i
r
e
c
t
i
o
n
a
l
t
r
a
n
sf
o
r
m
e
r
s
f
o
r
l
a
n
g
u
a
g
e
u
n
d
e
r
s
t
a
n
d
i
n
g
,
”
Pro
c
e
e
d
i
n
g
s
o
f
N
AA
C
L
-
H
L
T
2
0
1
9
,
2
0
1
9
,
p
p
.
4
1
7
1
–
4
1
8
6
.
[
2
4
]
S
.
S
i
d
h
u
,
S
.
S
.
K
h
u
r
a
n
a
,
M
.
K
u
mar,
P
.
S
i
n
g
h
,
a
n
d
S
.
S
.
B
a
m
b
e
r
,
“
S
e
n
t
i
m
e
n
t
a
n
a
l
y
si
s
o
f
H
i
n
d
i
l
a
n
g
u
a
g
e
t
e
x
t
:
a
c
r
i
t
i
c
a
l
r
e
v
i
e
w
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
3
,
n
o
.
1
7
,
p
p
.
5
1
3
6
7
–
5
1
3
9
6
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
0
4
2
-
0
2
3
-
1
7
5
3
7
-
6.
[
2
5
]
V
.
Y
a
d
a
v
,
P
.
V
e
r
ma
,
a
n
d
V
.
K
a
t
i
y
a
r
,
“
L
o
n
g
sh
o
r
t
t
e
r
m
m
e
m
o
r
y
(
LST
M
)
m
o
d
e
l
f
o
r
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
i
n
so
c
i
a
l
d
a
t
a
f
o
r
e
-
c
o
m
merc
e
p
r
o
d
u
c
t
s
r
e
v
i
e
w
s
i
n
H
i
n
d
i
l
a
n
g
u
a
g
e
s
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
1
5
,
n
o
.
2
,
p
p
.
7
5
9
–
7
7
2
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s4
1
8
7
0
-
0
2
2
-
0
1
0
1
0
-
y.
[
2
6
]
M
.
G
r
e
e
s
h
m
a
a
n
d
P
.
S
i
mo
n
,
“
B
i
d
i
r
e
c
t
i
o
n
a
l
g
a
t
e
d
r
e
c
u
r
r
e
n
t
u
n
i
t
w
i
t
h
g
l
o
v
e
e
mb
e
d
d
i
n
g
a
n
d
a
t
t
e
n
t
i
o
n
me
c
h
a
n
i
sm
f
o
r
m
o
v
i
e
r
e
v
i
e
w
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
Pro
c
e
d
i
a
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
2
3
3
,
p
p
.
5
2
8
–
5
3
6
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
r
o
c
s.
2
0
2
4
.
0
3
.
2
4
2
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
J
y
o
ti
S
.
Ve
r
m
a
is
a
n
a
c
c
o
m
p
li
sh
e
d
a
c
a
d
e
m
ic
p
ro
fe
ss
io
n
a
l
with
a
stro
n
g
b
a
c
k
g
ro
u
n
d
i
n
c
o
m
p
u
ter
a
p
p
li
c
a
ti
o
n
s
a
n
d
re
se
a
rc
h
.
S
h
e
h
o
l
d
s
a
M
a
ste
r
o
f
Co
m
p
u
ter
Ap
p
li
c
a
ti
o
n
s
(
M
CA)
d
e
g
re
e
fr
o
m
Ve
e
r
Na
rm
a
d
S
o
u
th
G
u
jara
t
Un
iv
e
rsity
a
n
d
is
c
u
rre
n
tl
y
p
u
rsu
i
n
g
h
e
r
P
h
.
D.
fr
o
m
Ch
a
n
g
a
Ch
a
ru
sa
t
Un
iv
e
rsit
y
.
W
it
h
a
p
a
s
sio
n
f
o
r
e
d
u
c
a
ti
o
n
,
s
h
e
h
a
s
b
e
e
n
a
c
ti
v
e
l
y
in
v
o
l
v
e
d
in
lec
tu
rin
g
sin
c
e
2
0
0
6
,
sh
a
p
i
n
g
t
h
e
m
in
d
s
o
f
a
sp
iri
n
g
stu
d
e
n
ts
.
P
re
se
n
tl
y
,
sh
e
is
c
o
n
tri
b
u
ti
n
g
h
e
r
e
x
p
e
rti
se
a
s
a
fa
c
u
lt
y
m
e
m
b
e
r
a
t
Ra
jj
u
S
h
ro
ff
ROFE
L
Un
iv
e
rsity
,
wh
e
re
sh
e
c
o
n
ti
n
u
e
s
to
i
n
sp
ire
a
n
d
g
u
i
d
e
stu
d
e
n
t
s
in
t
h
e
ir
a
c
a
d
e
m
ic
a
n
d
p
ro
fe
ss
io
n
a
l
jo
u
rn
e
y
s.
S
h
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
j
y
o
t
i.
s.v
e
rm
a
a
@g
m
a
il
.
c
o
m
.
J
a
i
m
in
N.
Un
d
a
v
ia
wo
rk
i
n
g
a
s
a
n
As
so
c
iate
P
ro
fe
ss
o
r
in
S
m
t.
Ch
a
n
d
a
b
e
n
M
o
h
a
n
b
h
a
i
P
a
tel
In
sti
tu
te
o
f
C
o
m
p
u
ter
A
p
p
li
c
a
ti
o
n
s,
F
a
c
u
lt
y
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
Ap
p
li
c
a
ti
o
n
s,
Ch
a
ro
tar
Un
i
v
e
rsit
y
o
f
S
c
ien
c
e
a
n
d
Tec
h
n
o
l
o
g
y
,
C
h
a
n
g
a
.
He
g
o
t
h
is
d
o
c
to
ra
te
fro
m
Ch
a
ru
sa
t
Un
i
v
e
rsity
.
He
h
a
s
p
u
b
li
sh
e
d
1
9
in
ter
n
a
ti
o
n
a
l
p
a
p
e
rs
,
1
n
a
ti
o
n
a
l,
1
i
n
tern
a
ti
o
n
a
l
b
o
o
k
c
h
a
p
ter,
a
n
d
1
in
ter
n
a
ti
o
n
a
l
b
o
o
k
.
He
p
o
ss
e
ss
e
s
1
8
y
e
a
rs
o
f
e
x
te
n
siv
e
tea
c
h
i
n
g
e
x
p
e
rien
c
e
.
His
re
se
a
rc
h
a
re
a
is
b
ig
d
a
ta
a
n
a
ly
t
ics
,
ro
b
o
ti
c
s,
I
o
T,
a
n
d
m
a
c
h
in
e
lea
rn
i
n
g
.
He
is
se
rv
in
g
5
i
n
tern
a
ti
o
n
a
l
j
o
u
r
n
a
ls
a
s
a
re
v
iew
e
r
a
n
d
lo
o
k
in
g
f
o
r
m
o
r
e
a
d
v
a
n
c
e
m
e
n
ts
in
th
e
fiel
d
o
f
ro
b
o
t
ics
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
jaim
in
u
n
d
a
v
ia.m
c
a
@c
h
a
ru
sa
t.
a
c
.
in
.
Evaluation Warning : The document was created with Spire.PDF for Python.