T
E
L
KO
M
NI
K
A
,
V
ol
.
14,
N
o.
3,
S
ept
em
ber
20
16,
pp.
10
67
~
107
6
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
2671
10
67
R
ec
ei
v
ed
S
ept
e
mb
er
11
,
2
01
5
;
R
ev
i
s
ed
Ma
y
1
0
,
201
6
;
A
c
c
ept
e
d
M
ay
24
,
20
1
6
A
S
e
n
t
i
m
e
n
t
K
n
o
w
led
g
e Disc
o
v
er
y
M
o
d
el
in
T
w
it
t
er
’s
T
V
Co
n
t
e
n
t
U
si
n
g
S
t
o
ch
ast
ic
G
r
ad
ien
t
Desc
en
t
A
l
g
or
i
th
m
L
i
r
a
R
u
h
w
i
n
a
n
i
n
g
s
i
h
*
,
T
a
u
fi
k
D
j
a
tn
a
B
ogor
A
gr
i
c
u
l
t
ur
a
l
U
ni
v
er
s
i
t
y
,
B
ogor
,
I
ndon
es
i
a
J
l
.
R
ay
a
D
ar
m
ag
a
,
K
am
p
us
I
P
B
D
a
r
m
a
ga
,
B
o
gor
1
6680
.
P
h
one
:
+
62 2
51 8
622
642
*
C
or
r
es
po
ndi
ng a
ut
hor
,
e
-
m
ai
l
:
pur
pl
e_f
i
m
@
y
ahoo
.
c
o
m
A
b
st
r
act
E
x
pl
o
s
i
v
e us
age of
s
oc
i
al
m
ed
i
a c
an be a r
i
c
h s
o
ur
c
e f
or
da
t
a m
i
ni
ng.
M
eanw
hi
l
e,
i
n
c
r
ea
s
i
ng and
v
ar
i
a
t
i
on
of
T
V
s
how
s
or
T
V
pr
ogr
am
s
m
ot
i
v
at
e
p
eopl
e
t
o
w
r
i
t
e
c
om
m
ent
s
v
i
a
s
oc
i
al
m
edi
a.
S
oc
i
al
net
w
or
k
c
ont
ai
n
s
abu
nda
nt
i
nf
or
m
at
i
o
n
w
hi
c
h
i
s
un
s
t
r
u
c
t
ur
ed,
h
et
er
o
geneo
us
,
hi
g
h
di
m
en
s
i
o
nal
a
n
d
a
l
ot
o
f
noi
s
e.
A
bunda
nt
dat
a c
an be a r
i
c
h s
our
c
e of
i
nf
or
m
at
i
on but
i
t
i
s
di
f
f
i
c
u
l
t
t
o be m
anua
l
l
y
i
dent
i
f
i
e
d
.
T
hi
s
r
es
ear
c
h
pr
opo
s
ed an a
ppr
oa
c
h t
o per
f
or
m
pr
epr
oc
es
s
i
ng of
u
ns
t
r
u
c
t
ur
e
d,
noi
s
y
an
d v
er
y
di
v
er
s
ed dat
a;
t
o f
i
nd
pat
t
er
n
s
of
t
he i
n
f
or
m
at
i
o
n a
nd k
now
l
ed
ge of
u
s
er
a
c
t
i
v
i
t
i
es
on
s
oc
i
al
m
ed
i
a i
n f
or
m
of
pos
i
t
i
v
e an
d
negat
i
v
e
s
ent
i
m
ent
on
T
w
i
t
t
er
T
V
c
ont
ent
.
S
om
e
m
et
hodol
ogi
e
s
and
t
e
c
hn
i
que
s
w
er
e
u
s
ed
t
o
per
f
or
m
pr
epr
o
c
es
s
i
ng.
T
h
er
e
w
er
e
r
e
m
ov
i
ng
p
unc
t
uat
i
on
a
nd
s
y
m
b
ol
s
,
r
em
o
v
i
n
g
num
ber
,
r
epl
ac
i
ng
num
ber
s
i
nt
o
l
et
t
er
s
,
t
r
a
ns
l
at
i
on
of
A
l
a
y
w
or
ds
,
r
em
ov
i
ng
s
t
op
w
or
d,
and
S
t
em
m
i
ng
us
i
ng
P
or
t
er
A
l
gor
i
t
hm
.
T
he
m
et
hodol
og
y
t
o u
s
ed
f
i
nd
pat
t
er
ns
o
f
i
nf
or
m
at
i
o
n and
k
now
l
edge o
f
s
o
c
i
al
m
edi
a i
n t
hi
s
s
t
udy
i
s
S
t
o
c
ha
s
t
i
c
G
r
adi
ent
D
es
c
en
t
(
S
G
D
)
.
T
he
t
ex
t
pr
e
pr
o
c
es
s
i
n
g pr
odu
c
ed
a m
or
e
s
t
r
uc
t
u
r
ed
t
ex
t
,
r
e
d
uc
ed
no
i
s
e
an
d
r
educ
ed
t
h
e
di
v
er
s
i
t
y
of
t
e
x
t
.
T
hus
,
pr
epr
oc
e
s
s
i
ng af
f
e
c
t
e
d
t
o
t
he ac
c
ur
a
c
y
and
p
r
oc
e
s
s
i
ng
t
i
m
e.
T
he
ex
per
i
m
ent
r
e
s
u
l
t
s
s
how
e
d
t
h
a
t
t
he
u
s
e
of
S
G
D
f
or
d
i
s
c
ov
er
y
of
t
he
po
s
i
t
i
v
e
and
nega
t
i
v
e
s
ent
i
m
ent
t
ent
t
o
be f
a
s
t
er
f
or
l
ar
g
e an
d s
t
r
ea
m
dat
a.
T
he P
er
c
ent
a
ge of
m
ax
i
m
um
ac
c
ur
a
c
y
to
f
in
d
s
ent
i
m
ent
pat
t
er
ns
i
s
88%
.
Ke
y
w
o
rd
s
:
s
t
oc
ha
s
t
i
c
gr
adi
e
nt
des
c
e
nt
,
o
pi
ni
on m
i
n
i
ng
,
s
ent
i
m
ent
anal
y
s
i
s
,
s
t
em
m
i
ng
por
t
er
,
s
t
r
eam
d
at
a
m
i
ni
ng
C
o
p
y
r
i
g
h
t
©
20
16 U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1
.
I
n
tr
o
d
u
c
ti
o
n
T
he
v
ar
i
et
y
of
T
V
P
r
ogr
am
w
as
pr
es
ent
e
d t
o t
he au
di
e
nc
es
c
an be good q
ua
l
i
t
y
c
ont
e
nt
,
t
ouc
hi
ng,
m
a
k
i
ng peo
pl
e anno
y
e
d,
m
onot
ono
us
an
d s
o f
or
t
h.
P
eop
l
e m
a
y
c
om
m
ent
on t
he
T
w
i
t
t
er
T
V
c
ont
e
nt
.
T
w
i
t
t
er
T
V
c
ont
ent
i
s
th
e
t
w
i
t
t
er
ac
c
ount
t
h
at
us
e
d
b
y
T
V
pr
o
g
r
am
pr
ov
i
d
er
s
as
a
m
edi
a
t
o
c
om
m
en
t
on
t
he
i
r
T
V
s
ho
w
s
,
ex
am
pl
e:
@
K
i
k
and
y
s
h
o
w
,
@
m
at
an
aj
w
a
an
d
ot
her
.
T
hes
e c
om
m
ent
s
c
an be
us
ef
ul
i
nf
or
m
at
i
on f
o
r
T
V
c
ont
ent
pr
o
v
i
der
or
bec
om
e a r
i
c
h
s
our
c
e of
i
nf
or
m
at
i
on f
or
da
t
a m
i
ni
ng.
B
i
l
l
i
ons
of
i
nf
or
m
at
i
o
n c
an
be
t
ak
en f
r
o
m
s
oc
i
al
m
edi
a
w
eb
pages
.
T
ex
t
c
om
m
ent
s
ar
e
goo
d i
n
di
c
a
t
or
s
of
us
er
s
at
i
s
f
ac
t
i
on.
S
en
t
i
m
ent
an
al
y
s
i
s
al
g
or
i
t
hm
s
of
f
er
an
anal
y
s
i
s
of
t
h
e
us
e
r
s
’
pr
ef
er
enc
es
,
i
n
w
h
i
c
h
t
h
e
c
om
m
ent
s
m
a
y
no
t
be
a
s
s
oc
i
at
ed
w
i
t
h
an ex
p
l
i
c
i
t
r
at
i
ng.
T
hus
,
t
h
i
s
ana
l
y
s
i
s
w
i
l
l
al
s
o
ha
v
e a
n i
m
pac
t
on t
he p
opu
l
ar
i
t
y
of
a gi
v
e
n m
edi
a
s
ho
w
.
A
ut
om
at
i
c
T
V
c
ont
e
nt
an
al
y
s
i
s
i
s
v
er
y
i
m
por
t
ant
f
or
t
he
ef
f
i
c
i
ent
i
ndex
i
ng
and
r
e
t
r
i
e
v
a
l
of
t
he
v
as
t
am
ount
of
T
V
c
ont
ent
a
v
ai
l
a
bl
e
not
on
l
y
f
or
end
us
er
s
s
o
t
hat
t
he
y
c
a
n
eas
i
l
y
ac
c
es
s
c
ont
ent
of
i
nt
er
es
t
,
b
ut
al
s
o
f
or
c
ont
ent
pr
oduc
er
s
and
br
oadc
as
t
er
s
i
n or
d
er
t
o i
de
nt
i
f
y
c
o
p
y
r
i
gh
t
v
i
ol
a
t
i
o
ns
and t
o f
i
l
t
er
or
c
har
ac
t
er
i
z
e un
des
i
r
a
bl
e c
ont
ent
.
T
he i
nf
or
m
at
i
ons
w
er
e c
apt
ur
ed f
r
om
s
oc
i
al
m
edi
a
pr
oduc
e
g
ol
d
us
ef
ul
i
nf
or
m
at
i
on
f
or
t
he
or
gani
z
at
i
o
n.
Mi
ni
ng
t
ho
us
a
nds
of
v
i
e
w
i
ng
c
hoi
c
es
and m
i
l
l
i
ons
of
pat
t
er
ns
,
adv
er
t
i
s
er
s
and T
V
net
w
or
k
s
i
dent
i
f
y
ho
us
eho
l
d
c
har
ac
t
er
i
s
t
i
c
s
,
t
as
t
es
,
an
d
d
e
s
ir
e
t
o c
r
eat
e an
d d
el
i
v
er
c
us
t
om
t
ar
get
ed a
dv
er
t
i
s
i
ng [
1
-
3
].
T
he
bac
k
gr
ound of
t
hi
s
r
es
ear
c
h i
s
t
o pr
o
v
i
d
e i
nf
or
m
at
i
on abo
ut
t
h
e pos
i
t
i
v
e or
ne
gat
i
v
e
s
ent
i
m
ent
t
o
t
he T
V
v
i
e
w
er
and c
ont
e
nt
pr
o
v
i
der
.
B
es
i
d
e,
T
V
c
ont
ent
pr
ov
i
der
s
ne
ed t
o do i
m
pr
ov
em
ent
i
n ev
er
y
ep
i
s
ode s
o
t
hat
t
he c
ont
ent
us
ef
ul
qu
al
i
f
i
ed and
gi
v
e
a pos
i
t
i
v
e
v
a
l
ue t
o t
he s
oc
i
et
y
.
T
V
c
ont
e
nt
t
ha
t
t
ak
i
ng a
l
ot
of
audi
enc
e
nee
ds
c
o
nt
i
n
uous
i
m
pr
ov
em
ent
t
o
av
o
i
d m
onot
o
n
y
an
d l
os
e
aud
i
enc
es
.
T
he
pr
oc
es
s
ed dat
a i
s
r
e
al
t
i
m
e and
s
t
r
eam
dat
a f
r
o
m
t
w
i
t
t
er
s
oc
i
al
m
edi
a t
hat
ev
er
y
t
i
m
e c
an be
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
67
–
1
076
1068
k
now
n
b
y
t
he
us
er
.
T
her
e
i
s
a
nu
m
ber
of
qual
i
t
y
at
t
r
i
b
ut
es
f
r
o
m
t
he
T
V
v
i
e
w
er
v
i
e
w
poi
nt
t
her
e
ar
e:
r
el
ev
anc
e
,
c
o
v
er
age,
d
i
v
er
s
i
t
y
,
un
der
s
t
an
dab
i
l
i
t
y
,
nov
el
t
y
,
and
s
er
end
i
p
i
t
y
[
4
].
T
he pr
obl
em
of
t
hi
s
r
es
ear
c
h
co
n
si
st
s
of
t
w
o.
T
he f
i
r
s
t
,
s
oc
i
al
m
edi
a da
t
a
c
ont
ai
n
abun
dan
t
i
nf
or
m
at
i
on
w
hi
c
h
i
s
uns
t
r
uc
t
ur
ed,
het
er
oge
n
eous
,
h
i
g
h d
i
m
ens
i
ona
l
an
d
a l
ot
of
no
i
s
e.
S
ec
on
d,
A
bun
dant
dat
a c
a
n be a r
i
c
h s
our
c
e of
i
nf
or
m
at
i
on,
but
i
t
i
s
di
f
f
i
c
ul
t
t
o i
dent
i
f
y
m
anua
l
l
y
and
ho
w
t
o
f
i
nd
pat
t
er
ns
of
i
nf
or
m
at
i
on an
d k
no
w
l
e
dg
e
of
s
oc
i
al
m
edi
a
us
er
ac
t
i
v
i
t
i
es
i
n
t
he
f
or
m
of
pos
i
t
i
v
e
and
ne
gat
i
v
e
s
ent
i
m
ent
on
t
w
i
t
t
er
T
V
c
ont
en
t
.
V
as
t
,
N
o
i
s
e,
d
i
s
t
r
i
but
ed and
uns
t
r
uc
t
ur
ed
ar
e
c
har
ac
t
e
r
i
s
t
i
c
s
of
s
oc
i
al
m
edi
a
da
t
a
[
5
]
.
P
r
epr
oc
es
s
i
ng
c
a
n
i
m
pr
ov
e
t
he
ac
c
ur
ac
y
and ef
f
i
c
i
enc
y
of
m
i
ni
ng al
gor
i
t
hm
s
i
nv
ol
v
i
n
g di
s
t
anc
e m
eas
ur
em
ent
s
.
F
eat
ur
e
s
el
ec
t
i
on t
ec
hn
i
q
ues
i
s
on
e of
t
he
m
os
t
i
m
por
t
ant
and f
r
e
quent
l
y
us
e
d i
n pr
e
p
r
oc
es
s
i
ng dat
a
m
i
ni
ng.
T
hi
s
t
ec
hn
i
q
ue r
e
d
uc
es
t
he
num
ber
of
f
eat
ur
es
t
hat
ar
e i
nv
ol
v
ed
i
n de
t
er
m
i
ni
ng a t
ar
get
c
l
as
s
v
a
l
ues
,
r
educ
i
ng
t
he
i
r
r
el
e
v
a
nt
f
eat
ur
e,
r
e
dund
ant
,
an
d
t
he
dat
a
t
hat
t
en
d
t
o
m
i
s
under
s
t
andi
n
g
of
t
he
t
a
r
get
c
l
as
s
t
hat
m
a
k
es
i
m
medi
at
e
ef
f
ec
t
f
or
appl
i
c
at
i
on
.
S
o
c
ia
l
m
e
d
ia
dat
a
i
s
a s
t
r
eam
dat
a t
hat
r
equi
r
es
m
et
hodo
l
og
i
es
a
nd al
gor
i
t
hm
s
c
an oper
at
e
w
i
t
h a l
i
m
i
t
ed
r
es
our
c
e bot
h i
n t
er
m
s
of
t
i
m
e
or
m
e
m
or
y
h
ar
d
w
ar
e.
Mor
eo
v
er
,
i
t
c
a
n han
dl
e
dat
a t
hat
c
han
ges
ov
er
t
i
m
e [
6
-
8]
.
T
he pr
opos
ed s
ol
ut
i
on t
o
t
he ab
ov
e pr
ob
l
em
s
is
b
y
m
a
k
i
ng s
om
e
m
odul
es
f
or
pr
epr
oc
es
s
i
ng
and
dat
a m
i
ni
ng m
odel
i
ng us
i
ng
S
G
D
m
et
hod.
T
w
e
et
d
at
a
w
as
t
ak
en us
i
ng
T
w
i
t
t
er
S
t
r
eam
i
ng A
P
I
pr
ov
i
de
d b
y
t
w
i
t
t
er
.
T
he d
at
a i
s
per
f
or
m
ed pr
epr
oc
es
s
i
ng
w
h
i
c
h
el
i
m
i
nat
es
pun
c
t
u
at
i
on a
nd s
y
m
bo
l
s
,
el
i
m
i
nat
i
n
g
num
ber
,
r
epl
ac
e n
um
ber
s
i
nt
o l
et
t
er
s
,
t
r
ans
l
at
i
o
n
of
A
l
a
y
w
or
ds
a
nd
el
i
m
i
nat
e
s
t
op
w
or
d.
T
he
nex
t
s
t
ep
i
s
per
f
or
m
i
ng
w
or
d
s
t
em
m
i
ng
us
i
ng
P
or
t
er
al
gor
i
t
hm
appr
oac
h
f
or
I
ndones
i
a
l
a
ngu
ag
e.
P
r
epr
oc
es
s
i
ng
i
s
an
i
m
por
t
ant
s
t
ep
t
o
t
he c
l
as
s
i
f
i
c
at
i
on pr
oc
es
s
and nec
es
s
ar
y
f
or
c
l
eans
i
n
g s
oc
i
al
m
edi
a dat
a t
h
at
f
i
l
l
ed
w
i
t
h no
i
s
e
and
uns
t
r
uc
t
ur
ed
s
o t
ha
t
t
h
e dat
a i
s
r
ea
d
y
t
o
be pr
oc
e
s
s
ed t
o t
h
e nex
t
s
t
e
p.
P
r
ep
r
oc
es
s
i
ng
w
a
s
m
ade
f
or
I
ndones
i
a l
ang
u
age an
d m
odul
es
of
pr
epr
oc
es
s
i
ng al
gor
i
t
h
m
c
r
eat
ed
b
y
t
he aut
hor
ex
c
ept
s
t
em
m
i
ng al
g
or
at
h
m
w
hi
c
h ad
opt
ed t
he
P
or
t
er
al
gor
i
t
hm
f
or
I
ndones
a l
a
ngu
age
.
C
l
as
s
i
f
i
c
at
i
on
al
gor
i
t
hm
w
as
us
ed
t
o f
i
nd
pat
t
er
ns
a
n
d i
nf
or
m
at
i
on
i
n
t
h
i
s
s
t
u
d
y
i
s
S
t
oc
h
as
t
i
c
G
r
adi
en
t
D
es
c
ent
.
S
G
D
is
s
ui
t
abl
e f
or
l
ar
ge and s
t
r
e
am
d
at
a
.
S
t
oc
h
as
t
i
c
G
r
adi
ent
D
es
c
ent
i
s
v
er
s
at
i
l
e t
ec
h
ni
ques
t
ha
t
h
av
e pr
o
v
e
n i
n
v
a
l
ua
bl
e
as
a l
ear
n
i
ng a
l
g
or
i
t
hm
f
or
l
ar
ge dat
as
et
s
.
F
r
o
m
t
he r
es
ear
c
h t
hat
ha
s
been
do
ne,
S
G
D
m
odel
gi
v
e ef
f
ec
t
t
o a s
h
or
t
pr
oc
es
s
i
ng t
i
m
e t
o
pr
oc
es
s
a
l
ot
of
dat
a
or
dat
a
s
t
r
eam
s
.
R
es
ear
c
h
about
k
no
w
l
e
dg
e
di
s
c
ov
er
y
on
T
w
i
t
t
er
s
t
r
eam
i
ng dat
a had be
en d
one b
y
B
i
f
et
and F
r
ank
.
B
a
s
ed on s
om
e
t
es
t
s
t
hat
hav
e been c
ar
r
i
e
d
out
,
S
G
D
m
odel
s
r
ec
om
m
ende
d f
or
t
he
dat
a
s
t
r
eam
w
i
t
h
t
he
det
er
m
i
nat
i
on
of
t
he
a
ppr
opr
i
at
e
l
ear
n
i
n
g r
at
e
[
8
-
9]
.
Pr
e
v
i
ou
s
r
es
ear
c
h on t
he
pos
i
t
i
v
e
and
neg
at
i
v
e
s
ent
i
m
ent
ha
s
been
don
e b
y
P
ut
r
a
nt
i
and
W
i
nar
k
o
[
10
]
,
nam
el
y
t
w
i
t
t
er
s
e
nt
i
m
ent
a
nal
y
s
i
s
f
or
I
ndo
nes
i
an
l
an
g
uage
w
i
t
h
t
h
e
Max
i
m
u
m
E
nt
r
op
y
a
nd S
u
ppor
t
V
ec
t
or
Mac
h
i
ne
.
R
e
s
ear
c
h on s
e
nt
i
m
ent
a
na
l
y
s
i
s
on T
w
i
t
t
e
r
dat
a
h
as
b
een
don
e
us
i
ng
k
er
nel
t
r
ee
a
nd
f
eat
ur
e
-
ba
s
ed
m
odel
s
[
11]
.
T
hen,
T
aboad
a
a
nd
h
i
s
t
eam
hav
e be
en
do
i
ng
r
es
ear
c
h
on Lex
i
c
on
-
B
as
ed
M
et
ho
ds
f
or
ex
t
r
ac
t
i
ng s
ent
i
m
ent
f
r
o
m
t
ex
t
[
12]
.
I
n a
ddi
t
i
o
n,
O
p
i
n
i
on m
i
ni
n
g an
d s
ent
i
m
ent
ana
l
y
s
i
s
on a T
w
i
t
t
er
d
at
a s
t
r
e
am
bas
ed o
n t
h
ei
r
em
ot
i
onal
c
ont
e
nt
as
pos
i
t
i
v
e,
neg
at
i
v
e
an
d i
r
r
e
l
e
v
an
t
b
y
G
ok
ul
ak
r
i
s
hnan
[
13]
.
T
her
e w
er
e t
w
o m
ai
n obj
ec
t
i
v
e
s
of
t
hi
s
s
t
ud
y
:
t
o p
er
f
or
m
pr
epr
oc
es
s
i
ng t
o a
ddr
es
s
uns
t
r
uc
t
ur
ed
d
at
a,
a
l
o
t
of
noi
s
e
a
nd
h
et
er
o
gen
eous
o
r
di
v
er
s
e;
f
i
n
d
pat
t
er
ns
of
i
nf
or
m
at
i
on
an
d
k
now
l
e
dge
of
s
oc
i
al
m
edi
a us
er
ac
t
i
v
i
t
i
es
i
n t
he f
or
m
o
f
pos
i
t
i
v
e a
nd n
ega
t
i
v
e s
ent
i
m
ent
o
n
t
w
i
t
t
er
T
V
c
ont
ent
.
2.
R
e
sea
r
ch
M
et
h
o
d
T
he c
as
e s
t
ud
y
i
n t
hi
s
r
es
ear
c
h
w
as
t
o
det
er
m
i
ne t
h
e pos
i
t
i
v
e
or
ne
gat
i
v
e s
ent
i
m
ent
bas
ed
on
t
w
e
et
of
T
V
c
on
t
ent
.
T
he t
w
e
et
d
at
a
t
ak
en
us
i
n
g T
w
i
t
t
er
S
t
r
e
am
i
ng
A
P
I
.
T
hen
i
t
i
s
t
ak
en c
ont
i
nu
ous
l
y
and s
t
or
ed i
n a t
ab
l
e i
n r
eal
t
i
m
e.
I
n t
he dat
a
bas
e,
t
he t
w
eet
s
t
hat
ha
v
e be
en
c
ol
l
ec
t
ed
at
t
h
e
t
abl
e,
pr
oc
es
s
ed
(
par
s
ed)
,
and
i
t
s
r
es
ul
t
di
s
t
r
i
but
e
d
t
o
s
ev
er
al
t
ab
l
es
f
or
pr
epar
at
i
o
n on
t
he
nex
t
pr
o
c
es
s
.
2.
1.
P
r
ep
r
o
ce
ssi
n
g
T
o ans
w
er
t
he pr
ob
l
em
of
uns
t
r
uc
t
ur
ed dat
a,
d
i
v
er
s
e and l
ot
s
of
noi
s
e
s
o
m
e
pr
epr
oc
es
s
i
ng
m
et
hods
a
nd t
ec
h
ni
ques
c
ar
r
i
e
d ou
t
.
T
he abo
v
e
ar
e t
h
e t
e
c
hni
q
ues
an
d
m
et
hodol
o
gi
es
w
er
e
us
e
d i
n
t
he
pr
epr
oc
es
s
i
ng
i
n t
h
i
s
r
es
ear
c
h.
I
n gen
er
al
,
t
he pr
epr
oc
es
s
i
ng
c
an be
des
c
r
i
be
d o
n F
i
gur
e
1.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
S
ent
i
me
nt
K
no
w
l
e
dg
e D
i
s
c
ov
er
y
M
od
el
i
n
T
w
i
t
t
er
’
s
T
V
C
o
nt
en
t
…
(
L
i
r
a R
u
hw
i
n
ani
ngs
i
h
)
1069
Remove
Number at the
Beginning
Word
Replace
Number with
Letter
Remove
Repeat Letter
Translate
Alay Word
Remove
Stop Word
Stemming
Remove
Punctuation
&
Symbol
F
i
gur
e
1
.
P
r
epr
oc
es
s
i
n
g s
t
a
ges
2.
1.
1
.
E
lim
in
a
t
e
A
ll P
u
n
c
tu
a
ti
o
n
a
n
d
S
y
m
b
o
l
s
T
hi
s
i
nf
or
m
at
i
on ge
ner
al
l
y
does
not
add t
o t
he
und
er
s
t
andi
ng of
t
ex
t
an
d
w
i
l
l
m
ak
e i
t
har
der
t
o
p
ar
s
e
t
he
w
or
ds
on
s
om
e
c
o
m
m
ent
s
on
T
w
i
t
t
er
.
T
hi
s
i
nf
or
m
at
i
on
i
nc
l
u
d
es
s
i
ng
l
e
and
doub
l
e
qu
ot
at
i
o
n
m
ar
k
s
,
par
ent
hes
es
,
pu
nc
t
uat
i
o
n,
an
d
ot
her
s
y
m
bol
s
s
uc
h
as
d
ol
l
ar
s
i
g
ns
a
nd
s
t
ar
s
.
F
or
ex
am
pl
es
,
"
H
ar
us
k
ah
k
i
t
a
me
mbay
ar
k
e
mba
l
i
neg
er
i
i
n
i
!
!
,
t
p
s
el
a
ma
k
i
t
a
t
er
us
ber
t
any
a
ap
a n
egar
a
i
ni
d
apat
me
mber
i
k
an.
"
C
hr
i
s
t
i
ne H
ak
i
m
#MN
,
a
nd af
t
er
t
he pr
oc
es
s
of
r
em
ov
i
ng
punc
t
u
at
i
on
an
d
s
y
m
bol
s
,
t
h
e
s
ent
e
nc
e
bec
om
es
"
H
ar
us
k
ah
k
i
t
a
m
em
bay
ar
k
em
ba
l
i
neger
i
i
n
i
,
t
p s
el
a
ma k
i
t
a t
er
us
ber
t
any
a ap
a neg
ar
a
i
ni
da
pat
me
mb
er
i
k
an.
"
C
hr
i
s
t
i
n
e H
ak
i
m
#MN
.
2
.1
.2
.
E
l
i
m
i
n
a
t
e
N
u
m
b
e
r
s
i
n
fr
o
n
t
o
f
th
e
W
o
r
d
s
F
or
ex
am
pl
es
,
"
2har
i
"
bec
om
es
"
har
i
"
.
J
us
t
l
i
k
e r
e
m
ov
e s
y
m
bol
,
r
em
ov
e
num
ber
al
s
o
us
e m
at
c
h c
har
ac
t
er
f
unc
t
i
on t
o
r
em
ov
e num
ber
b
y
r
e
gul
ar
ex
pr
es
s
i
on
of
.
net
l
i
br
ar
y
.
2
.1
.3
.
R
e
p
l
a
c
e
N
u
m
b
e
r
w
i
th
L
e
tt
e
r
T
he f
ol
l
o
w
i
ng
i
s
an
al
gor
i
t
h
m
f
or
r
epl
ac
e num
ber
w
i
t
h
l
et
t
er
:
1.
G
et
l
i
s
t
dat
a c
on
v
er
s
i
on f
r
om
dat
abas
e b
as
e o
n
T
abl
e
1
2.
I
ns
er
t
l
i
s
t
dat
a c
on
v
er
s
i
on t
o has
h
v
ar
i
ab
l
e
3.
C
hec
k
nu
m
er
i
c
c
har
ac
t
er
i
n t
he m
i
ddl
e
of
i
nput
s
t
r
i
n
g and c
hec
k
i
f
i
t
i
s
not
c
ont
ai
ned
al
l
num
er
i
c
c
har
ac
t
er
us
i
ng
r
e
gex
,
i
f
t
r
ue t
hen
:
C
hec
k
w
et
her
i
t
i
s
c
on
t
ai
n
ed s
t
r
i
ng
"
0
0"
,
i
f
t
r
u
e t
hen
r
epl
ac
e
i
t
b
y
"
u
"
el
s
e
i
f
i
t
i
s
c
ont
a
i
ne
d
s
t
r
i
ng t
h
at
m
at
c
h w
i
t
h
dat
a i
n T
abl
e 1
,
i
t
w
i
l
l
be r
e
pl
ac
e
d b
y
c
on
v
er
t
i
on
dat
a f
r
o
m
T
abl
e 1.
4.
I
f
poi
nt
c
i
s
f
al
s
e,
t
h
en r
et
ur
n t
he
or
i
g
i
n
of
i
np
ut
s
t
r
i
n
g.
T
abl
e
1
.
Li
s
t
of
c
onv
er
t
i
ng
t
he n
um
ber
s
i
nt
o
l
et
t
er
s
N
um
ber
C
onv
er
s
i
o
n
0
o
00
u
1
i
2
C
opy
c
har
bef
or
e “
2”
3
e
4
a
5
s
6
g
7
t
8
b
9
g
2.
1.
4
.
E
l
i
m
i
n
at
e R
ep
e
at
ed
l
et
t
e
r
s
E
x
am
pl
es
:
"
P
a
g
ii
ii
ii
ii
ii
i
"
bec
om
es
"
pagi
"
.
C
hec
k
i
f
i
t
ha
s
r
epeat
ed
l
et
t
er
t
h
en t
ak
e t
he f
i
r
s
t
c
har
ac
t
er
on
l
y
.
2.
1.
5
.
T
r
a
n
s
l
a
ti
n
g
th
e
“
A
l
ay
”
W
o
r
d
s
i
n
to
N
o
r
m
a
l
W
o
r
d
s
A
l
a
y
w
or
d
i
s
ex
c
es
s
i
v
e or
s
t
r
ange
w
or
ds
a
nd
al
s
o
us
e
w
r
i
t
i
n
g m
i
x
ed num
ber
s
at
o
nc
e or
us
i
ng
u
pp
er
c
as
e
and
l
o
w
er
c
as
e
l
et
t
er
s
t
h
at
ar
e
n
o
t
r
e
as
onab
l
e
[
14]
.
I
t
i
s
s
t
or
ed
i
n
t
he
d
at
a
bas
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
67
–
1
076
1070
and c
al
l
ed
as
A
l
a
y
da
t
a
d
i
c
t
i
o
nar
y
.
T
he pr
o
gr
am
w
i
l
l
c
hec
k
i
nt
o
dat
abas
e
w
h
e
t
her
t
he
w
or
d
i
nc
l
u
ded
t
he
A
l
a
y
w
or
ds
or
not
,
a
nd t
hen
t
r
ans
l
at
e t
he
w
or
ds
A
l
a
y
i
n
t
o n
or
m
al
w
or
ds
.
E
x
am
pl
es
:
"
e
M4N
k
4L
aY
G
hH
eT
oO
" b
e
c
om
es
"
emang
al
ay
gi
t
u
”
.
2.
1.
6
.
E
l
i
m
i
n
a
ti
n
g
th
e
S
to
p
W
o
r
d
S
t
op
w
or
d i
s
a
w
or
d
t
hat
ha
s
no m
eani
ng
and
do
n
ot
c
ont
r
i
b
ut
e
dur
i
ng
t
h
e pr
oc
es
s
i
ng
of
t
he
d
at
a.
S
t
op
w
or
d
i
s
a
l
s
o
s
t
or
ed
i
n
t
h
e
da
t
ab
as
e
da
t
a
as
da
t
a
s
t
o
p
w
or
ds
d
i
c
t
i
onar
y
ar
e
l
i
k
e
A
l
a
y
dat
a.
I
f
t
h
er
e
ar
e
w
or
ds
t
ha
t
ar
e
l
i
s
t
ed
i
n t
h
e
t
ab
l
e s
t
op
w
or
d,
t
he
w
or
d i
s
r
em
ov
e
d.
E
x
am
pl
es
:
"
J
i
k
a
di
a
per
g
i
"
bec
om
es
"
per
gi
”
.
2.
1.
5
.
N
e
x
t S
te
p
I
s
P
e
r
fo
r
m
th
e
W
o
r
d
S
te
m
m
i
n
g
S
t
em
m
i
ng
i
s
t
he
pr
oc
es
s
o
f
t
r
ans
f
or
m
at
i
on
of
w
or
ds
c
ont
a
i
ne
d i
n a
doc
um
ent
t
o
bas
i
c
w
or
d (
r
oot
w
or
d
)
w
i
t
h c
er
t
ai
n r
u
l
es
.
T
he al
g
or
i
t
hm
t
hat
us
ed t
o s
t
em
m
i
ng i
s
c
al
l
e
d A
l
gor
i
t
hm
P
or
t
er
f
or
t
he
I
nd
ones
i
a
l
an
guag
e [
1
5]
.
T
he f
ol
l
o
w
i
ng
i
s
P
or
t
er
al
g
or
i
t
hm
:
S
t
ep
1 :
R
em
ov
i
n
g t
h
e par
t
i
c
l
es
.
S
t
ep
2 :
R
em
ov
i
n
g pos
s
es
s
i
v
e pr
on
ou
n.
S
t
ep
3 :
D
e
l
et
e t
he f
i
r
s
t
pr
e
f
i
x
.
I
f
t
her
e i
s
a s
ec
ond pr
ef
i
x
,
go t
o s
t
e
p 4.
1
,
i
f
t
her
e i
s
s
uf
f
i
x
t
hen go
t
o s
t
ep
4.
2.
S
t
ep
4 :
S
t
ep
4.
1
R
em
ov
i
n
g s
ec
on
d
pr
ef
i
x
,
pr
oc
eed
t
o s
t
e
p 5.
1.
S
t
ep
4.
2 R
em
ov
i
ng s
uf
f
i
x
,
i
f
i
t
i
s
no
t
f
ound
t
he
n t
he
w
or
d i
s
as
s
um
ed t
o
be t
he r
o
ot
w
or
d
.
I
f
f
ound t
hen
pr
oc
ee
d t
o
s
t
e
p 5.
2f
S
t
ep
5 :
S
t
ep
5.
1
R
em
ov
i
n
g s
uf
f
i
x
,
t
hen t
he f
i
n
al
w
or
d
i
s
as
s
um
ed t
o
be
t
he r
o
ot
w
or
d.
S
t
ep
5.
2
D
e
l
et
i
ng
a s
ec
on
d
pr
ef
i
x
,
t
hen
t
he
f
i
na
l
w
or
d
i
s
as
s
u
m
ed t
o be
t
h
e r
oot
w
or
d.
2.
2.
C
l
a
s
s
i
fi
c
a
ti
o
n
M
o
d
e
l
U
s
i
n
g
S
to
c
h
a
s
t
i
n
c
G
r
ad
i
en
t
D
escen
t
(
S
G
D
)
T
he r
eas
on of
t
he us
e of
S
G
D
i
n t
hi
s
s
t
ud
y
bec
a
us
e t
he dat
a be
i
ng pr
oc
es
s
ed
i
s
a dat
a
s
t
r
eam
t
hat
i
s
c
ont
i
n
uous
l
y
updat
ed eac
h t
i
m
e.
T
he s
t
r
eam
dat
a has
c
har
ac
t
er
i
s
t
i
c
t
hat
f
l
o
w
s
c
ont
i
n
uous
l
y
,
v
er
y
l
ar
g
e an
d r
eal
t
i
m
e.
S
G
D
has
t
h
e a
bi
l
i
t
y
t
o us
e on
l
y
one
t
r
ai
ni
n
g s
a
m
pl
e f
r
o
m
t
r
ai
n
i
ng s
et
t
o do t
he upd
at
e f
or
a par
a
m
et
er
i
n a par
t
i
c
ul
ar
i
t
er
at
i
o
n.
I
n t
he pr
oc
e
s
s
i
ng of
s
t
r
ea
m
dat
a,
t
he
da
t
a do n
ot
w
a
i
t
unt
i
l
pi
l
l
i
n
g up b
ut
t
he d
at
a
pr
oc
es
s
i
ng i
s
don
e c
ont
i
n
uous
l
y
i
n r
e
al
t
i
m
e
as
l
on
g
as
t
he
dat
a
i
s
bei
n
g
s
t
r
eam
ed.
I
n
m
os
t
m
a
c
hi
ne
l
ear
n
i
ng,
t
he
obj
e
c
t
i
v
e
f
unc
t
i
on
i
s
of
t
en t
he
c
um
ul
at
i
v
e s
um
of
t
he er
r
or
o
v
er
t
he
t
r
ai
ni
ng ex
am
pl
es
.
B
ut
t
he s
i
z
e
of
t
he t
r
a
i
n
i
ng
ex
am
pl
es
s
et
m
i
ght
be v
er
y
l
ar
ge
and h
enc
e c
om
put
i
ng t
he ac
t
ua
l
gr
ad
i
ent
w
ou
l
d
be
c
o
m
put
at
i
ona
l
l
y
ex
p
ens
i
v
e.
I
n S
G
D
m
et
hod,
c
om
put
e an es
t
i
m
at
e or
appr
ox
i
m
at
i
on t
o t
h
e
di
r
ec
t
i
on m
ov
e
onl
y
on t
hi
s
appr
ox
i
m
at
i
on
.
I
t
i
s
c
al
l
e
d as
s
t
oc
has
t
i
c
bec
a
us
e t
h
e ap
pr
ox
i
m
at
e
di
r
ec
t
i
on t
ha
t
i
s
c
o
m
put
ed at
ev
er
y
s
t
e
p c
an be t
ho
u
ght
of
a r
andom
v
a
r
i
ab
l
e
of
a s
t
oc
has
t
i
c
pr
oc
es
s
.
T
her
e
ar
e t
w
o m
ai
n par
t
s
i
n
m
ac
hi
ne l
e
ar
ni
ng
nam
el
y
,
t
r
ai
ni
n
g
and t
e
s
t
i
ng.
T
he
dat
a
m
odel
gen
er
at
e
d f
r
o
m
t
he t
r
ai
n
i
n
g pr
oc
es
s
i
s
us
e
d
f
or
t
he pr
ed
i
c
t
i
o
n pr
oc
es
s
i
n s
e
nt
i
m
ent
di
s
c
ov
er
y
.
T
hi
s
i
s
an a
ns
w
e
r
t
o t
he
s
ec
ond
pr
ob
l
em
s
in
t
h
is
r
es
ear
c
h.
F
i
r
s
t
of
al
l
,
t
he t
w
eet
d
at
a t
hat
ha
v
e be
en l
abe
l
e
d pos
i
t
i
v
e (
y
es
)
and ne
gat
i
v
e (
no)
s
t
or
ed
i
n
A
ttr
i
b
u
te
-
R
e
l
at
i
o
n
F
i
l
e
F
or
m
at
(
A
R
F
F
)
f
i
l
e.
T
he
r
eas
on
f
or
us
i
ng
t
he
A
R
F
F
f
i
l
e
bec
aus
e
S
G
D
l
i
br
ar
y
t
hat
i
s
us
e
d c
om
es
f
r
o
m
t
he
W
e
k
a appl
i
c
a
t
i
o
n.
A
n A
R
F
F
f
i
l
e
i
s
a
n A
S
C
I
I
t
ex
t
f
i
l
e t
hat
des
c
r
i
bes
a l
i
s
t
of
i
ns
t
anc
e
s
s
har
i
ng a s
et
of
at
t
r
i
but
e
s
.
A
R
F
F
f
i
l
es
hav
e t
w
o di
s
t
i
nc
t
s
ec
t
i
ons
.
T
he
f
i
r
s
t
s
ec
t
i
on i
s
t
he H
ea
der
i
nf
or
m
at
i
on,
w
h
i
c
h i
s
f
ol
l
o
w
e
d t
he da
t
a i
nf
or
m
at
i
on.
T
he header
of
t
he
A
R
F
F
f
i
l
e
c
o
nt
a
i
ns
t
he
nam
e of
t
he r
e
l
at
i
o
n,
a
l
i
s
t
of
t
he
at
t
r
i
but
es
or
t
he c
o
l
u
m
n
s
i
n t
he
dat
a
.
A
n ex
am
pl
e
hea
der
o
n t
h
e
dat
as
et
l
o
ok
s
l
i
k
e t
hi
s
:
@
r
el
at
i
on '
M
y
A
r
f
f
T
V
C
ont
en
t
'
@
at
t
r
i
but
e t
ex
t
s
t
r
i
ng
@
at
t
r
i
but
e c
l
as
s
{
no
,
y
es
}
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
S
ent
i
me
nt
K
no
w
l
e
dg
e D
i
s
c
ov
er
y
M
od
el
i
n
T
w
i
t
t
er
’
s
T
V
C
o
nt
en
t
…
(
L
i
r
a R
u
hw
i
n
ani
ngs
i
h
)
1071
Raw Data
Labelled
(
after
preprocessing
)
Create Arff
file
Setting
Parameter
Tokenizer
Data
Dictionary
Update
Classifier
using SGD
Data
Modeling
Prediction
Testing
Training
Arrf for
Testing
Arrf for
Learning
Data
Prediction
F
i
gur
e
2
.
F
l
o
w
c
l
as
s
i
f
i
c
at
i
o
n m
odel
T
he
D
at
a
of
t
he
A
R
F
F
f
i
l
e
l
ook
s
l
i
k
e t
he f
ol
l
o
w
i
n
g:
@
dat
a
'
S
em
al
a
m n
ang
i
s
K
i
c
k
A
n
dy
S
how
'
,
no
'
k
i
c
k
andy
s
how
ey
uy
a
nt
k
ar
t
i
k
ar
ahm
aeu
l
i
s
aon
gt
s
t
ar
g
u
mi
l
ar
l
ama
t
y
g t
a
ng '
,
no
'
K
i
c
k
A
n
dy
S
h
ow
k
au
j
o
dohk
uk
ena
p
ak
ar
na
j
o
doh
ng
at
ur
j
od
oh
at
ur
J
omb
l
o
MahG
i
t
u'
,
y
es
'
K
i
c
k
A
n
dy
S
h
ow
o
m s
ay
a s
u
l
i
a
t
r
eg
i
s
t
l
og'
,
n
o
A
R
F
F
dat
a
i
s
di
v
i
de
d i
n
t
o
t
w
o
par
t
s
,
o
ne
f
or
t
r
ai
n
i
ng
and
a
not
h
er
one f
or
t
es
t
i
n
g.
T
he
pr
oc
es
s
f
l
ow
i
s
as
s
um
ed f
or
t
he
e
v
a
l
ua
t
i
o
n m
odel
b
as
ed
on t
he
pr
es
en
t
a
t
i
on
of
t
r
ai
n
i
ng
an
d
t
es
t
i
n
g.
T
he det
er
m
i
nat
i
o
n
of
t
he v
al
ue of
a pr
es
ent
at
i
on
is
bas
e
d on us
er
i
nput
.
T
her
e ar
e s
o
m
e
par
am
et
er
s
i
n t
he t
r
ai
n
i
ng
pr
oc
es
s
,
ex
a
m
pl
es
l
ear
n
i
n
g r
at
e,
epoc
h,
l
am
bda and
et
c
.
T
he nex
t
pr
oc
es
s
i
s
t
he
t
ok
eni
z
er
.
I
n
l
ex
i
c
al
ana
l
y
s
i
s
,
to
k
eni
z
at
i
on
i
s
t
he
pr
oc
es
s
of
br
eak
i
ng
a
s
t
r
eam
of
t
ex
t
up
i
nt
o
w
or
ds
,
phr
as
es
,
s
y
m
bol
s
,
or
ot
h
er
m
eani
ng
f
ul
el
em
ent
s
c
al
l
ed
t
ok
ens
.
S
G
D
i
s
t
he
m
et
hodol
og
y
u
s
ed
t
o
f
i
nd
pat
t
er
ns
of
i
nf
or
m
at
i
on
i
n
t
he
f
or
m
of
pos
i
t
i
v
e
and
nega
t
i
v
e
s
ent
i
m
ent
on
t
w
i
t
t
er
T
V
c
ont
ent
on
t
h
i
s
r
es
e
ar
c
h.
S
t
oc
h
as
t
i
c
G
r
adi
ent
D
es
c
ent
(
S
G
D
)
i
s
an a
l
gor
i
t
hm
t
hat
i
s
us
ed f
or
l
ar
ge
-
s
c
al
e
l
ear
ni
n
g pr
ob
l
em
s
.
S
G
D
has
an ex
c
el
l
ent
per
f
or
m
anc
e i
n s
ol
v
i
ng
l
ar
g
e
-
s
c
al
e pr
ob
l
em
s
.
S
G
D
ef
f
i
c
i
ent
i
n p
er
f
or
m
i
ng c
l
as
s
i
f
i
c
at
i
o
n e
v
en
i
f
i
t
i
s
bas
ed
on
non
-
di
f
f
er
ent
i
a
bl
e
l
os
s
f
unc
t
i
o
n (
F
(θ
)
)
[4
].
(
)
=
2
|
|
|
|
2
+
∑
[
1
−
(
+
)
]
w
i
s
t
he
w
e
i
g
ht
v
ec
t
or
,
b
i
s
t
he
bi
as
,
i
s
a
r
egul
at
i
on
par
am
et
er
,
and
t
he
c
l
as
s
l
abe
l
y
as
s
u
m
ed
{
+
1,
-
1}
.
E
x
p
l
i
c
i
t
s
pec
i
f
i
c
at
i
o
n
of
l
ear
n
i
ng
r
at
e
i
s
an
i
m
por
t
ant
t
hi
n
g
t
ha
t
r
e
l
at
es
t
o
t
he
t
i
m
e c
hange
i
n t
he
dat
a s
t
r
eam
.
B
i
f
et
a
nd F
r
a
nk
hav
e
i
n
v
es
t
i
ga
t
ed
r
es
ear
c
h
on t
he T
w
i
t
t
er
s
t
r
eam
s
of
dat
a
us
i
ng
s
e
v
e
r
al
m
et
hods
:
Mu
l
t
i
nom
i
al
N
ai
v
e
B
a
y
es
,
S
t
oc
has
t
i
c
G
r
a
di
e
nt
D
es
c
ent
and
H
oef
f
di
ng
T
r
ee.
B
as
ed
on t
he
t
es
t
s
t
hat
ha
v
e
be
en
do
ne
of
t
he
t
hr
e
e m
et
hods
,
m
odel
s
S
G
D
w
i
t
h
t
he
appr
opr
i
a
t
e
l
ear
n
i
n
g
r
at
e
i
s
r
ec
om
m
ended
t
o
pr
oc
es
s
t
he
d
at
a
s
t
r
e
am
.
T
he
f
or
m
ul
at
i
o
n
of
updat
e c
l
as
s
i
f
i
er
i
n
S
G
D
i
s
(
F
ne
w
)
[
8]
:
=
∑
∑
.
(
1
−
(
.
)
)
=
0
+
.
|
(
+
)
|
=
0
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
67
–
1
076
1072
Input
:
Instance
,
bool
updateDictionary
,
m
_
bias
Class index
<
0
(
negative number if it
'
s
undefined
)
Tokenizer
wx
=
Dotprod
Check if label
instance
(
class
attribute
) ==
0
Y
=
-
1
Y
Y
=
1
N
Z
=
y
* (
wx
+
m
_
bias
)
;
double multiplier
=
1
;
Check if instance data
count
(
rows
) ==
0
multiplier
=
1
-
(
m
_
learningRate
*
m
_
lambda
) /
m
_
t
;
Y
multiplier
=
1
-
(
m
_
learningRate
*
m
_
lambda
) /
instance data count
;
N
Update weight with multiplier
,
By looping for all weight from dictionary
:
weight
(
index
) *=
multiplier
;
N
End
Y
Start
F
i
gur
e 3.
F
l
o
w
u
pd
at
e c
l
as
s
i
f
i
er
par
t
1 r
ef
er
enc
ed f
r
om
W
e
k
a l
i
br
ar
i
es
W
h
er
e:
n =
num
ber
of
dat
a t
w
eet
r
o
w
s
i
=
s
um
of
i
t
er
at
i
on
w
=
w
ei
ght
=
L
ear
n
i
ng
R
at
e
λ
=
Lam
bda
y
=
S
t
at
us
of
s
ent
i
m
ent
(
y
e
s
or
no)
=
ac
c
um
ul
at
i
on
of
w
e
i
g
ht
(
per
w
or
d)
b =
bi
as
.
Check if z
<
1
dLoss
=
1
Y
dLoss
=
0
N
double factor
=
m
_
learningRate
*
y
*
dloss
;
Input
_
map has more entry
M
_
bias
+=
factor
N
m
_
t
++
;
Output
:
Hashmap dictionary
Get word from entry element
;
Get count word from entry element
;
Y
Check if weight is not null
Result
+=
count
word
*
factor
Y
N
Start
End
F
i
gur
e
4.
F
l
o
w
u
pd
at
e c
l
as
s
i
f
i
er
par
t
2 r
ef
er
enc
ed f
r
om
W
e
k
a l
i
br
ar
i
es
I
ns
t
anc
e
dat
a
i
s
s
t
or
e
d i
n
has
hm
ap di
c
t
i
onar
y
,
w
hi
c
h
i
s
a s
t
r
uc
t
ur
e
of
dat
a
bas
ed o
n
has
hi
ng,
and al
l
o
w
s
s
t
o
r
in
g
t
he
obj
ec
t
as
k
e
y
v
al
ue
pai
r
.
I
n t
h
i
s
c
as
e H
as
hM
ap (
k
e
y
,
obj
ec
t
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
S
ent
i
me
nt
K
no
w
l
e
dg
e D
i
s
c
ov
er
y
M
od
el
i
n
T
w
i
t
t
er
’
s
T
V
C
o
nt
en
t
…
(
L
i
r
a R
u
hw
i
n
ani
ngs
i
h
)
1073
v
a
l
ue)
m
eans
t
hat
eac
h
r
ec
or
d
dat
a
has
a
k
e
y
w
or
d
a
nd
an
obj
ec
t
v
al
ue.
B
es
i
de
t
hat
,
i
t
s
obj
ec
t
v
a
l
ue i
s
an o
bj
ec
t
t
hat
c
ont
ai
n
ed t
w
o d
at
a v
al
ues
,
w
h
i
c
h ar
e num
ber
of
w
or
d an
d
w
e
i
gh
t
.
I
f
dat
a
i
ns
t
anc
e
i
s
gr
eat
er
t
han
z
e
r
o,
t
hen
d
at
a
r
a
w
pos
i
t
i
o
n
t
o
be
p
l
ac
ed
a
t
r
and
om
pos
i
t
i
on.
T
he
goa
l
i
s
dat
a i
nput
c
an b
e c
l
os
er
t
o t
he
r
eal
c
ond
i
t
i
on
i
n
i
m
pl
e
m
ent
at
i
on
w
i
t
h
di
f
f
er
ent
i
np
ut
s
.
I
n t
h
e f
unc
t
i
on
U
pd
at
e
c
l
a
s
s
i
f
i
er
t
her
e
ar
e s
e
v
er
al
s
ub
-
pr
oc
es
s
es
,
n
a
m
el
y
"
t
ok
eni
z
er
"
and "
dot
pr
od
"
.
T
ok
eni
z
er
(
t
ok
eni
z
at
i
o
n)
has
a m
ai
n pr
oc
es
s
f
or
br
ea
k
i
ng a s
t
r
ea
m
of
t
ex
t
up i
nt
o
w
or
ds
,
phr
as
es
,
or
ot
her
m
ean
i
ngf
ul
el
em
ent
s
c
al
l
ed
t
ok
ens
.
I
f
c
l
as
s
i
ndex
i
s
gr
e
at
er
t
ha
n
z
er
o,
t
hen
nex
t
pr
oc
es
s
i
s
t
ok
eni
z
er
.
I
np
ut
d
at
a
of
t
ok
eni
z
er
i
s
a
n
i
ns
t
anc
e
d
at
a,
an
d
i
n
t
hi
s
c
as
e
,
a
n
i
ns
t
anc
e
dat
a i
s
on
e t
w
e
et
dat
a,
t
hat
r
e
pr
es
ent
ed i
n
has
hm
ap.
T
o
k
eni
z
er
has
s
om
e c
hec
k
i
ng
pr
oc
es
s
es
t
o v
al
i
d
at
e d
at
a
,
bef
or
e goi
ng t
o
i
t
s
m
ai
n pr
oc
es
s
.
F
i
r
s
t
,
i
t
v
a
l
i
dat
es
i
np
ut
i
ns
t
anc
e
dat
a,
t
o c
h
ec
k
i
f
i
t
i
s
N
U
LL
or
not
,
and
s
ec
ond
c
hec
k
i
ng i
f
i
t
s
at
t
r
i
bu
t
e t
y
p
e i
s
a
s
t
r
i
ng d
at
a t
y
p
e.
3.
R
e
su
l
t
s an
d
A
n
al
y
s
i
s
T
he
pur
pos
e
of
t
hi
s
ex
per
i
m
ent
w
as
t
o
de
t
er
m
i
ne
t
he
ac
c
ur
ac
y
of
t
he
S
G
D
m
et
hod
b
y
c
hang
i
ng s
om
e par
am
et
er
s
.
Mor
eo
v
er
,
t
o k
no
w
ho
w
t
o i
nf
l
uenc
e
s
om
e di
f
f
er
enc
es
pr
epr
oc
es
s
i
ng
pr
oc
es
s
aga
i
ns
t
c
l
as
s
i
f
i
c
at
i
o
n r
es
ul
t
s
i
n
ac
c
ur
ac
y
and
pr
oc
es
s
i
ng
t
i
m
e.
T
ot
al
dat
a
i
s
us
ed f
or
t
es
t
i
ng an
d l
ea
r
ni
ng
is
745 t
w
e
et
s
.
A
n
d
her
e ar
e
s
om
e of
t
he
ex
per
i
m
ent
s
w
er
e
c
onduc
t
e
d
.
3.
1.
E
v
a
l
u
a
ti
o
n
M
o
d
e
l
u
s
i
n
g
S
p
l
i
t T
e
s
t
3.
1.
1
.
P
r
ep
r
o
ce
ssi
n
g
o
n
t
h
e C
l
a
ssi
f
i
cat
i
o
n
A
c
cu
r
ac
y an
d
P
r
o
c
essi
n
g
T
i
m
e
T
he
f
i
r
s
t
t
es
t
i
s
t
he
ana
l
y
s
i
s
of
pr
epr
oc
es
s
i
ng
on
t
he
c
l
as
s
i
f
i
c
at
i
o
n
ac
c
ur
a
c
y
a
nd
pr
oc
es
s
i
ng t
i
m
e.
D
at
a i
s
d
i
v
i
ded i
nt
o t
w
o par
t
s
,
90%
t
r
ai
n
i
ng d
at
a a
nd 10
%
t
es
t
i
ng d
at
a.
T
he
gr
aph
be
l
o
w
s
ho
w
s
t
h
e
v
ar
i
at
i
on
of
t
he
pr
epr
oc
es
s
i
n
g
of
t
he
c
l
as
s
i
f
i
c
at
i
on
ac
c
ur
ac
y
w
hi
c
h
t
e
nd
t
o r
i
s
e,
w
hen t
h
e dat
a has
t
hr
oug
h s
ev
er
a
l
s
t
ages
of
pr
epr
oc
es
s
i
ng
.
C
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed i
ns
t
a
nc
e
i
s
ac
c
ur
ac
y
i
n c
l
as
s
i
f
y
i
n
g t
h
e dat
a
.
S
i
gn
i
f
i
c
ant
i
nc
r
eas
e
s
oc
c
ur
r
ed f
or
dat
a t
hat
has
been t
hr
oug
h
pr
epr
oc
es
s
i
ng
s
t
ep;
t
hos
e
ar
e r
em
ov
e s
t
op
w
o
r
d
and
t
he s
t
em
m
i
ng.
S
t
ep
R
em
ov
e s
t
o
p
w
or
d
m
eans
t
o el
i
m
i
nat
e w
or
ds
t
hat
i
nc
l
ude a S
t
op
w
or
d
i
n t
he dat
a t
w
eet
,
bec
au
s
e i
t
does
no
t
c
ont
r
i
bu
t
e t
o pr
oc
es
s
of
c
l
as
s
i
f
i
c
at
i
on.
W
hi
l
e t
he
s
t
em
m
i
ng w
i
l
l
m
a
k
e w
or
ds
of
t
w
eet
o
n
ly
co
n
si
st
t
he r
o
ot
w
or
d
.
T
he s
uf
f
i
x
,
pr
ef
i
x
par
t
i
c
l
e,
s
uf
f
i
x
w
i
l
l
be
e
l
i
m
i
nat
ed
i
n
t
he
pr
oc
es
s
of
s
t
em
m
i
ng.
T
he
t
i
m
e
r
equi
r
ed
pr
oc
es
s
t
en
d
s
t
o
dec
r
eas
e
i
f
t
he
dat
a
h
as
bee
n
t
hr
o
ugh
i
n
s
ev
er
al
pr
epr
oc
es
s
i
n
g
s
t
eps
,
s
ho
w
n
i
n
gr
ap
h as
i
l
l
us
t
r
at
ed
i
n F
i
gur
e 5.
F
i
gur
e
5.
C
har
t
ex
ec
u
t
i
o
n
t
i
m
e on pr
epr
oc
es
s
i
ng
v
ar
i
at
i
ons
T
he nex
t
t
es
t
i
ng us
ed l
ear
n
i
ng r
at
e
=
0.
01
,
epoc
h =
1
0
0,
w
i
t
h
v
ar
i
a
t
i
ons
of
l
ear
ni
n
g dat
a
,
and
al
s
o
dat
a
t
es
t
i
ng.
D
at
a
t
es
t
i
n
g
9
0%
u
nt
i
l
1
0%
us
i
n
g
dec
r
em
ent
s
t
ep
10%
and
D
at
a
t
r
ai
ni
n
g
10%
unt
i
l
90%
,
a
nd
us
i
ng
i
nc
r
em
ent
s
t
ep 1
0%
.
T
he d
at
a i
s
d
i
v
i
d
ed
i
n
t
o t
r
ai
ni
n
g a
nd
t
es
t
i
n
g
dat
a
.
I
n t
h
e f
i
r
s
t
ex
p
er
i
m
ent
,
t
h
e t
r
ai
ni
ng
dat
a
ar
e s
et
b
y
10%
of
t
he
t
ot
al
dat
a
and
da
t
a
t
es
t
i
ng
b
y
90%
of
t
he
t
o
t
al
dat
a.
F
ur
t
her
t
e
s
t
i
ng
b
y
ad
di
n
g
t
he
dat
a
t
r
ai
n
i
ng
an
d
r
ed
uc
e
t
he
d
at
a
t
es
t
i
n
g.
F
ig
ur
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
67
–
1
076
1074
6
s
ho
w
s
t
hat
t
he
m
or
e
t
r
ai
ni
n
g
d
at
a,
t
he
c
l
as
s
i
f
i
c
at
i
o
n
ac
c
ur
ac
y
w
i
l
l
t
en
d
t
o
r
i
s
e
.
Mor
e
t
r
ai
ni
ng
dat
a
w
i
l
l
r
e
qu
i
r
e i
nc
r
eas
i
n
g
t
he ex
ec
u
t
i
o
n
t
im
e
;
i
t
i
s
s
ho
w
n i
n F
i
g
ur
e 7
.
F
i
gur
e
6.
C
or
r
ec
t
l
y
c
l
as
s
i
f
i
e
d i
ns
t
anc
es
c
har
t
F
i
gur
e
7.
T
ot
al
t
i
m
e
ex
ec
ut
i
on C
h
ar
t
3.
1.
2
.
C
h
a
n
g
i
n
g
L
e
a
r
n
i
n
g
R
a
te
T
he
nex
t
ex
per
i
m
ent
is
c
ha
ngi
ng
l
ear
n
i
ng
r
a
t
e,
w
i
t
h
e
p
oc
h
100
f
or
eac
h
v
ar
i
at
i
on
of
t
he
am
ount
of
t
r
ai
n
i
n
g a
nd t
es
t
i
ng.
F
i
gur
e
8
s
ho
w
s
t
he
v
ar
i
at
i
o
ns
on
l
ear
n
i
ng
r
at
e
an
d
t
he
am
ount
of
t
r
ai
n
i
ng
d
at
a
.
L
ear
n
i
ng r
at
e 0.
00
1 gi
v
es
t
he c
l
as
s
i
f
i
c
at
i
o
n ac
c
ur
ac
y
v
a
l
u
es
w
hi
c
h t
end t
o r
i
s
e.
H
o
w
e
v
er
,
t
h
e ef
f
ec
t
o
f
c
hangi
n
g l
e
ar
ni
ng r
at
e a
ga
i
ns
t
t
he ex
ec
ut
i
on t
i
m
e i
s
not
t
oo
s
i
gni
f
i
c
ant
.
I
n
add
i
t
i
on t
hat
i
l
l
us
t
r
a
t
ed
i
n
F
i
gur
e 9
,
t
h
e
l
ear
n
i
n
g r
at
e
i
s
get
t
i
ng
s
m
al
l
er
,
s
o t
h
e pr
oc
es
s
w
i
l
l
t
ak
e a
l
on
g t
i
m
e t
en
d ex
ec
ut
i
on
w
hen t
he
num
ber
of
t
r
ai
ni
ng
dat
a
i
nc
r
eas
es
.
F
i
gur
e
8.
C
har
t
of
ep
oc
h c
h
anges
t
o c
or
r
ec
t
l
y
c
l
as
s
i
f
i
e
d i
ns
t
anc
es
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
S
ent
i
me
nt
K
no
w
l
e
dg
e D
i
s
c
ov
er
y
M
od
el
i
n
T
w
i
t
t
er
’
s
T
V
C
o
nt
en
t
…
(
L
i
r
a R
u
hw
i
n
ani
ngs
i
h
)
1075
3.
1.
3
.
E
p
o
c
h
C
h
a
n
g
e
s
to
C
o
r
r
e
c
tl
y
C
l
a
s
s
i
fi
e
d
I
n
s
ta
n
c
e
s
F
i
gur
e
8
s
ho
w
s
t
hat
c
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed
I
ns
t
anc
es
i
s
h
i
ghes
t
dur
i
ng
hi
gh
t
r
ai
ni
n
g s
et
a
nd
hi
g
h ep
oc
h
.
3.
2.
E
v
a
l
u
a
ti
o
n
M
o
d
e
l
u
s
i
n
g
C
r
o
s
s
v
a
l
i
d
a
ti
o
n
T
es
t
i
ng b
y
us
i
n
g c
r
os
s
v
al
i
d
at
i
on
w
i
t
h f
ol
ds
2,
3,
4,
5,
6,
7,
8,
9,
and 10 ob
t
ai
n
e
d
c
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed i
ns
t
anc
e
s
t
end t
o b
e s
t
ab
l
e
at
ar
o
un
d 84%
i
t
i
s
s
een
i
n
t
he
f
ol
l
o
w
i
ng
c
har
t
F
i
gur
e
9
s
ho
w
s
m
eas
ur
em
ent
of
c
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed
i
ns
t
anc
e
us
i
n
g
c
r
os
s
v
al
i
da
t
i
on
w
i
t
h
v
ar
i
at
i
ons
of
f
ol
ds
num
ber
.
C
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed
i
ns
t
an
c
e
l
ook
s
t
o
be
ar
ound
84
%
w
h
at
e
v
er
t
h
e
num
ber
of
f
ol
ds
.
F
i
gur
e
9.
C
har
t
c
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed i
ns
t
anc
es
us
i
ng
C
r
os
s
V
al
i
d
at
i
on
F
r
o
m
t
he
F
i
gur
e
10
be
l
o
w
s
how
s
t
h
at
t
h
er
e ar
e t
w
o
c
l
as
s
es
ar
e c
onc
ent
r
at
e
d
i
n t
w
o
pl
ac
es
.
T
op
r
i
ght
i
s
a
c
l
as
s
"
Y
es
"
t
he
p
os
i
t
i
v
e
s
ent
i
m
e
nt
.
W
hi
l
e
on
t
he
b
ot
t
om
l
ef
t
i
s
a
c
ol
l
ec
t
i
on
of
c
l
as
s
"
N
o
"
i
s
n
eg
at
i
v
e s
e
nt
i
m
ent
.
I
n F
i
g
ur
e 1
0
s
ho
w
s
t
hat
s
am
pl
e p
at
t
er
n
ou
t
pu
t
c
apt
ur
e
d t
h
a
t
des
c
r
i
be h
o
w
s
en
t
i
m
ent
p
oi
nt
di
s
t
r
i
but
i
on f
or
eac
h
w
or
d i
s
p
l
ac
ed
.
R
e
d dot
r
epr
es
ent
ed f
or
pos
i
t
i
v
e s
ent
i
m
ent
w
or
ds
,
and ot
her
w
i
s
e,
b
l
ue
dot
d
e
s
c
r
i
be f
or
negat
i
v
e s
ent
i
m
ent
w
or
d
.
I
f
i
t
s
pos
i
t
i
on
i
s
c
l
os
er
t
o t
i
p,
s
o i
t
i
s
s
i
g
ni
f
i
es
t
h
at
s
ent
i
m
ent
i
s
get
t
i
ng s
t
r
o
nger
a
nd i
t
m
ak
es
i
t
s
ac
c
ur
ac
y
b
ec
om
es
hi
gh.
A
s
t
he
des
c
r
i
pt
i
on
of
f
i
nd
i
ng
s
and
c
om
put
at
i
o
n
a
bo
v
e,
t
he
ad
v
ant
age
s
of
t
hi
s
s
t
ud
y
r
e
v
e
al
ed t
he
s
t
r
engt
h of
i
n
f
or
m
at
i
on h
i
di
ng f
r
o
m
T
V
audi
enc
e
’
s
T
w
i
t
t
er
s
.
O
ur
pr
epr
oc
es
s
i
ng
ap
pr
oac
h h
as
s
uc
c
eeded t
o di
c
o
v
er
ed
i
nf
or
m
at
i
on f
r
o
m
uns
t
r
uc
t
ur
ed and
no
i
s
y
dat
a.
A
s
s
o f
ar
,
t
he w
ho
l
e pr
oc
es
s
i
ng t
i
m
e t
end
ed
t
o s
hor
t
er
t
han pr
e
v
i
ous
s
t
udi
es
ab
out
s
ent
i
m
ent
ana
l
y
s
i
s
on t
he
s
t
r
ea
m
dat
a us
i
ng
Mu
l
t
i
n
om
i
al
N
ai
v
e B
a
y
es
and H
oef
f
di
ng T
r
ee.
S
om
e di
s
ad
v
a
nt
ag
es
i
nc
l
ud
e c
ur
r
ent
pr
ob
l
em
i
n ac
c
ur
a
c
y
as
w
e
l
l
as
t
h
e p
ar
t
i
a
l
r
e
p
r
es
ent
at
i
on
of
ho
w
t
h
e ev
al
uat
i
on
of
pub
l
i
c
m
edi
a s
uc
h
as
T
V
c
ont
ent
s
c
i
ent
i
f
i
c
al
l
y
c
om
put
ed.
B
as
ed o
n t
h
e
r
es
ul
t
s
of
pr
ev
i
o
us
s
t
udi
es
[
3]
ac
c
ur
ac
y
Mu
l
t
i
nom
i
al
N
a
i
v
e
B
a
y
es
m
et
hods
ar
e
bet
t
er
t
han
S
G
D
.
F
i
gur
e
10.
S
e
nt
i
m
ent
pa
t
t
er
n
4
.
C
o
n
c
l
u
s
i
o
n
T
he
dat
a
us
e
d
i
n
t
hi
s
s
t
ud
y
i
s
der
i
v
ed
f
r
om
t
he
dat
a
t
w
i
t
t
er
T
V
c
ont
ent
s
uc
h
as
c
o
m
m
ent
s
or
us
er
t
w
e
et
.
T
w
i
t
t
er
ac
c
ou
nt
w
h
i
c
h
bec
am
e t
he obj
ec
t
of
t
hi
s
s
t
ud
y
i
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
67
–
1
076
1076
@
k
i
c
k
and
y
s
ho
w
.
T
w
eet
da
t
a
t
hat
has
be
en
t
hr
o
ugh
pr
epr
oc
es
s
i
ng
c
an
b
e
m
or
e
s
t
r
uc
t
ur
ed
t
ex
t
,
r
educ
i
n
g n
oi
s
e
an
d r
ed
uc
i
n
g t
h
e d
i
v
er
s
i
t
y
.
P
r
epr
oc
es
s
i
ng ac
c
ur
ac
y
i
n e
l
i
m
i
nat
i
ng
al
l
pu
nc
t
uat
i
o
n
and
s
y
m
bol
s
,
el
i
m
i
nat
i
n
g num
ber
s
,
r
epl
ac
e num
ber
s
w
i
t
h l
et
t
er
s
,
e
l
i
m
i
nat
i
ng r
epeat
ed
l
et
t
er
r
eac
hes
100 %
w
i
t
h t
he e
x
ec
ut
i
on t
i
m
e of
eac
h appr
ox
i
m
at
el
y
2 m
i
l
l
i
s
ec
o
nds
f
or
760 t
w
e
et
s
.
W
h
er
eas
,
t
he ex
ec
ut
i
o
n t
i
m
e f
or
t
r
ans
l
at
i
ng t
he s
t
op
w
or
d,
el
i
m
i
nat
i
n
g s
t
op
w
or
ds
and s
t
em
m
i
ng
i
s
77,
111
and
12
30
m
i
l
l
i
s
ec
onds
r
es
pec
t
i
v
el
y
.
P
r
e
pr
oc
es
s
i
ng
s
t
a
ges
af
f
ec
t
t
o
t
he
hi
gher
ac
c
ur
ac
y
of
t
he c
l
as
s
i
f
i
c
at
i
on an
d r
educ
e pr
oc
es
s
i
ng
t
i
m
e.
T
hi
s
s
t
ud
y
s
uc
c
es
s
f
ul
l
y
ex
t
r
ac
t
e
d
pat
t
er
ns
of
i
nf
or
m
at
i
on a
nd
k
now
l
edg
e of
s
oc
i
a
l
m
edi
a
us
er
ac
t
i
v
i
t
i
es
i
n
t
h
e f
or
m
of
pos
i
t
i
v
e
an
d
nega
t
i
v
e
s
ent
i
m
ent
on
t
w
i
t
t
er
T
V
c
ont
e
nt
.
T
he r
es
ul
t
s
of
t
he
ex
per
i
m
ent
s
h
o
w
ed t
hat
l
ar
ge
am
ount
of
t
r
ai
n
i
n
g
d
at
a
c
a
n
af
f
ec
t
t
o
t
h
e
ac
c
ur
ac
y
of
t
he
c
l
as
s
i
f
i
c
at
i
o
n
of
p
os
i
t
i
v
e
an
d
neg
at
i
v
e
s
ent
i
m
ent
but
r
equ
i
r
e
a
l
o
n
ger
t
i
m
e
f
or
t
he
t
r
ai
ni
ng
pr
oc
es
s
.
Lear
n
i
ng
r
at
e
c
h
an
ges
l
i
t
t
l
e
ef
f
ec
t
on dur
at
i
o
n of
t
he pr
oc
es
s
and t
h
e ac
c
ur
ac
y
of
c
l
as
s
i
f
i
c
at
i
o
n.
Le
ar
ni
ng r
at
e
i
s
get
t
i
ng s
m
al
l
er
t
han
t
he
ex
ec
ut
i
on
t
i
m
e c
an i
nc
r
eas
e b
y
us
i
ng
l
ar
g
e t
r
ai
ni
ng d
at
a.
P
er
c
ent
ag
e of
c
or
r
ec
t
l
y
c
l
as
s
i
f
i
ed i
ns
t
anc
e
is
w
i
t
h
a
m
ax
i
m
u
m
of
88%
.
I
n
or
d
er
t
o
f
ul
f
i
l
l
a
n
i
de
al
s
t
ud
y
,
i
n
t
he
f
ut
ur
e
i
t
i
s
r
ec
o
m
m
ended
t
o
s
e
ar
c
h
f
or
a
hi
g
her
ac
c
ur
ac
y
l
ev
el
e
d a
l
gor
i
t
hm
t
hat
c
o
v
er
t
he
w
ho
l
e pr
oc
es
s
i
n T
V
c
ont
e
nt
e
v
a
l
u
at
i
on
s
c
i
ent
i
f
i
c
al
l
y
.
R
ef
er
en
ces
[1
]
P
el
ej
a
F
,
D
i
a
s
P
,
M
ar
t
i
n
s
F
.
A
r
ec
o
m
m
end
er
s
y
s
t
em
f
or
t
he T
V
on
t
he
w
eb:
i
nt
egr
at
i
n
g unr
at
ed
r
ev
i
ew
s
and m
ov
i
e r
at
i
ng
s
.
M
ul
t
i
m
edi
a S
y
s
t
em
s
.
201
3;
19
(
6)
:
1
-
16.
[2
]
H
an J
,
K
am
ber
M
,
P
ei
J
.
D
at
a M
i
ni
ng C
on
c
ept
s
a
nd T
ec
h
ni
que
s
.
S
ec
o
nd
E
d
i
t
i
on.
S
an
F
r
anc
i
s
c
o:
M
or
gan K
auf
m
a
n.
2
012.
[3
]
S
pangl
er
W
E
,
G
al
-
O
r
M,
Ma
y
J
H
.
U
s
i
ng
D
at
a
M
i
ni
ng t
o
P
r
of
i
l
e
T
V
V
i
ew
er
s
.
C
om
m
uni
c
a
t
i
on o
f
th
e
AC
M
.
P
i
t
t
s
bur
gh.
2003
;
46(
12)
:
67
-
7
2.
[4
]
B
am
bi
ni
R
,
C
r
em
ona
s
i
P
,
T
u
r
i
i
n
R
.
T
V
C
ont
en
t
A
n
al
y
s
i
s
:
R
ec
om
end
er
S
y
s
t
e
m
f
or
I
nt
er
ac
t
i
v
e
T
V
.
B
oc
a R
at
on:
C
R
C
P
r
e
s
s
.
201
2
.
[5
]
G
undec
ha
P
,
H
uan
L.
M
i
ni
ng
S
oc
i
a
l
M
edi
a:
A
B
r
i
ef
I
nt
r
odu
c
t
i
on.
I
n
:
S
m
i
th
J
C,
G
r
eenb
er
g
HJ
.
E
di
t
or
s
.
T
ut
or
i
al
s
i
n O
p
er
at
i
on
s
R
es
ear
c
h
:
N
ew
D
i
r
ec
t
i
on
s
i
n
I
n
f
or
m
at
i
c
s
,
O
pt
i
m
i
z
at
i
on,
Lo
gi
s
t
i
c
s
,
an
d
P
r
oduc
t
i
on
.
E
i
g
ht
h E
d
i
t
i
on.
A
r
i
z
ona:
I
NF
O
RM
S
;
2012:
1
-
17
.
[6
]
H
an J
,
K
am
ber
M
,
P
ei
J
.
D
at
a M
i
ni
ng C
on
c
ept
s
a
nd T
ec
h
ni
que
s
.
S
ec
o
nd
E
d
i
t
i
on.
S
an
F
r
anc
i
s
c
o:
M
or
gan K
auf
m
a
n.
2
012.
[7
]
D
j
at
na
T
,
Y
as
uhi
k
o
M
.
P
em
bandi
n
gan
s
t
abi
l
i
t
as
al
g
or
i
t
m
a
s
el
e
k
s
i
f
i
t
ur
m
engg
una
k
an
t
r
a
ns
f
or
m
a
s
i
r
ank
i
ng n
or
m
al
.
J
ur
n
al
I
l
m
u K
o
m
put
er
.
20
08;
6(
2)
:
1
-
6.
[8
]
Bi
f
e
t
A,
F
ra
n
k
E.
S
ent
i
m
ent
K
now
l
ed
ge
D
i
s
c
o
v
er
y
i
n
T
w
i
t
t
e
r
S
t
r
eam
i
ng
D
at
a
.
P
r
oc
e
edi
ngs
of
t
he
13
t
h
I
nt
er
na
t
i
o
nal
C
o
nf
er
e
nc
e,
D
S
2010.
C
anber
r
a.
2
010;
633
2:
1
-
15.
[9
]
B
ot
t
ou L.
S
t
o
c
ha
s
t
i
c
G
r
adi
en
t
D
es
c
en
t
T
r
i
c
k
.
I
n:
M
ont
av
on G
,
O
r
r
G
,
M
ul
l
er
K
.
Ed
i
t
o
rs
.
N
e
u
r
al
N
e
t
wo
r
k
:
T
r
i
c
k
of
t
he T
r
ade
.
S
ec
on
d E
di
t
i
on
.
H
ei
del
b
er
g:
S
pr
i
nger
B
er
l
i
n H
ei
del
b
er
g;
2012
:
421
-
4
36
.
[
10]
P
ut
r
ant
i
N
,
W
i
nar
k
o
E
.
2
014.
A
nal
i
s
i
s
S
ent
i
m
en
T
w
i
t
t
er
un
t
uk
T
ek
s
B
er
bah
as
a
I
nd
one
s
i
a de
nga
n
M
a
x
i
m
um
E
nt
r
opy
da
n S
up
por
t
V
ec
t
or
M
ac
hi
ne
.
J
CCS
.
8:
91
-
100
.
[
11]
A
gr
aw
al
A
,
X
i
e B
,
V
ov
s
ha
I
,
R
am
bow
O
,
P
as
s
on
nea
u R
.
S
ent
i
m
ent
A
na
l
y
s
i
s
F
or
T
w
i
t
t
e
r
D
at
a
.
L
S
M
'
11 P
r
oc
e
edi
ngs
of
t
he
W
o
r
k
s
h
op on
Lan
guag
es
i
n S
o
c
i
al
M
edi
a.
S
t
r
oud
s
bur
g.
2
011:
30
-
38
.
[
12]
T
aboada
M
,
B
r
ook
e
J
,
T
of
i
l
o
s
k
i
M
,
V
ol
l
K
,
S
t
e
de
M
.
Lex
i
c
on
-
B
as
ed
M
et
hod
s
f
or
S
ent
i
m
en
t
A
nal
y
s
i
s
.
M
I
T
P
r
es
s
J
o
ur
nal
.
20
11;
37(
2
)
:
267
-
307.
[
13]
G
ok
ul
a
k
r
i
s
hn
an
B
,
P
r
i
y
ant
ha
n
P
,
R
agav
an
T
,
P
r
as
at
h
N
,
P
er
er
a
A
.
O
pi
ni
on
M
i
ni
ng
and
S
ent
i
m
ent
A
nal
y
s
i
s
o
n
a
T
w
i
t
t
er
D
at
a
S
t
r
eam
.
I
E
E
E
A
dv
anc
es
i
n
I
C
T
f
or
E
m
er
gen
i
ng
R
egi
o
ns
(
I
C
T
er
)
I
nt
er
na
t
i
o
nal
C
o
nf
er
e
nc
e
.
20
1
2;
10:
182
-
188.
[
14]
M
e
y
k
e.
P
enggu
naan
K
o
s
a
K
at
a
A
l
ay
O
l
eh
R
em
aj
a
P
ada
F
a
c
eboo
k
di
K
ot
a
B
e
ngk
ul
u.
M
as
t
er
T
hes
i
s
.
B
engk
ul
u:
P
os
t
gr
ad
uat
e
B
eng
k
ul
u
U
ni
v
er
s
i
t
y
;
201
3.
[
15]
A
gus
t
a L.
C
om
par
i
s
on
of
P
or
t
er
S
t
em
m
i
ng
A
l
gor
i
t
hm
and
N
a
z
i
ef
&
A
dr
i
ani
’
s
A
l
gor
i
t
hm
f
or
S
t
em
m
i
ng
I
ndon
es
i
an T
ex
t
D
o
c
um
ent
s
.
N
at
i
ona
l
C
o
nf
er
e
nc
e
an
d
I
nf
o
r
m
at
i
on
S
y
s
t
em
s
2009
.
B
a
l
i
.
2009;
03
6:
196
-
20
1.
Evaluation Warning : The document was created with Spire.PDF for Python.