T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
1
8
,
No.
3
,
J
une
2020
,
pp.
1
422
~
1
432
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v1
8
i
3
.
14791
1422
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
C
le
ve
r
e
e
:
an
a
r
t
ifi
c
ia
ll
y i
n
t
e
ll
ig
e
n
t
w
e
b
se
r
vi
c
e
f
or
J
ac
o
b
voi
c
e
c
h
at
b
o
t
Oct
avan
y,
Ar
ya
Wicaks
an
a
D
ep
ar
t
men
t
o
f
In
f
o
rmat
i
cs
,
U
n
i
v
ers
i
t
a
s
Mu
l
t
i
me
d
i
a
N
u
s
an
t
ara,
In
d
o
n
es
i
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Aug
3
,
2019
R
e
vis
e
d
J
a
n
2
0
,
2020
Ac
c
e
pted
F
e
b
26
,
2020
J
aco
b
i
s
a
v
o
i
ce
ch
at
b
o
t
t
h
a
t
u
s
e
W
i
t
.
ai
t
o
g
e
t
t
h
e
co
n
t
e
x
t
o
f
t
h
e
q
u
e
s
t
i
o
n
an
d
g
i
v
e
an
an
s
w
er
b
a
s
ed
o
n
t
h
at
co
n
t
e
x
t
.
H
o
w
ev
er,
J
ac
o
b
h
as
n
o
v
ari
a
t
i
o
n
i
n
an
s
w
er
an
d
co
u
l
d
n
o
t
reco
g
n
i
ze
t
h
e
co
n
t
e
x
t
w
e
l
l
i
f
i
t
h
as
n
o
t
b
een
l
earn
e
d
p
rev
i
o
u
s
l
y
b
y
t
h
e
W
i
t
.
ai
.
T
h
u
s
,
t
h
i
s
p
a
p
er
p
ro
p
o
s
es
t
w
o
feat
u
res
o
f
art
i
f
i
ci
a
l
i
n
t
el
l
i
g
en
ce
(A
I)
b
u
i
l
t
as
a
w
eb
s
erv
i
ce
:
t
h
e
p
ara
p
h
ra
s
e
o
f
an
s
w
er
s
u
s
i
n
g
t
h
e
St
ac
k
ed
Res
i
d
u
al
L
ST
M
mo
d
el
an
d
t
h
e
q
u
e
s
t
i
o
n
s
u
mmar
i
zat
i
o
n
u
s
i
n
g
Co
s
i
n
e
Si
m
i
l
ar
i
t
y
w
i
t
h
p
re
-
t
ra
i
n
e
d
W
o
rd
2
V
ec
a
n
d
T
ex
t
Ran
k
al
g
o
ri
t
h
m.
T
h
e
s
e
t
w
o
fea
t
u
re
s
are
n
o
v
e
l
d
es
i
g
n
s
t
h
a
t
are
t
ai
l
o
r
ed
t
o
J
aco
b
,
t
h
i
s
A
I
mo
d
u
l
e
i
s
cal
l
e
d
Cl
e
v
eree.
T
h
e
ev
a
l
u
a
t
i
o
n
o
f
Cl
ev
eree
i
s
carri
ed
o
u
t
u
s
i
n
g
t
h
e
t
ech
n
o
l
o
g
y
acce
p
t
a
n
ce
m
o
d
e
l
(T
A
M)
met
h
o
d
a
n
d
i
n
t
erv
i
ew
w
i
t
h
J
aco
b
ad
mi
n
s
.
T
h
e
res
u
l
t
s
s
h
o
w
t
h
a
t
7
9
.
1
7
%
o
f
res
p
o
n
d
e
n
t
s
s
t
r
o
n
g
l
y
a
g
ree
t
h
at
b
o
t
h
feat
u
re
s
are
u
s
ef
u
l
an
d
7
2
.
5
7
%
o
f
res
p
o
n
d
e
n
t
s
s
t
r
o
n
g
l
y
ag
ree
t
h
at
b
o
t
h
feat
u
re
s
are
eas
y
t
o
u
s
e.
K
e
y
w
o
r
d
s
:
C
os
ine
s
im
il
a
r
it
y
L
S
T
M
S
tac
ke
d
r
e
s
idual
T
e
xtR
a
nk
W
e
b
s
e
r
vice
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Ar
ya
W
ica
ks
a
na
,
De
pa
r
tm
e
nt
of
I
nf
or
mat
ics
,
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
,
S
c
ientia
B
ouleva
r
d
S
t.
,
Ga
ding
S
e
r
pong
,
T
a
nge
r
a
ng
-
15810,
B
a
nten,
I
ndone
s
ia.
E
mail:
a
r
ya
.
wic
a
ks
a
na
@umn.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
A
c
ha
tbot
is
a
mes
s
a
ging
pr
ogr
a
m
that
int
e
r
a
c
ts
with
us
e
r
s
li
ke
c
ha
tt
ing
with
pe
ople.
T
he
r
e
is
r
e
s
e
a
r
c
h
on
de
ve
lopi
ng
a
dialogue
s
ys
tem
or
c
h
a
tbot
us
ing
na
tur
a
l
langua
ge
that
is
us
e
f
ul
f
or
e
duc
a
ti
on,
c
us
tom
e
r
s
e
r
vice
,
a
nd
e
nter
tainment
pur
pos
e
s
[
1
-
3
]
.
I
n
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
(
UM
N)
,
ther
e
is
a
we
b
-
ba
s
e
d
voice
c
ha
tbot
c
a
ll
e
d
J
a
c
ob
whic
h
is
d
e
ve
loped
by
W
ij
a
ya
in
[
4
]
.
J
a
c
ob
us
e
s
the
W
it
.
a
i
platf
or
m
to
ge
t
the
c
ontext
o
f
the
us
e
r
’
s
que
s
ti
on
by
e
xtr
a
c
ti
ng
the
int
e
nt
(
the
goa
l
of
the
us
e
r
is
c
omi
ng
to
the
c
ha
tbot
)
a
nd
e
nti
ti
e
s
(
im
por
tant
va
r
iable
in
int
e
nt
that
he
l
ps
a
dd
r
e
leva
nc
e
to
a
n
int
e
nt)
of
the
que
s
ti
on.
T
he
r
e
f
or
e
,
the
c
ha
tbot
c
ould
r
e
p
ly
to
the
que
s
ti
on
ba
s
e
d
on
the
c
ontext
(
int
e
nt
a
nd
e
nti
ti
e
s
)
.
How
e
ve
r
,
J
a
c
ob
ha
s
two
s
hor
tcomings
whe
n
int
e
r
a
c
ti
ng
with
the
us
e
r
.
F
ir
s
t,
J
a
c
ob
only
r
e
pli
e
s
with
a
ns
we
r
s
that
a
r
e
a
lr
e
a
dy
pr
ogr
a
mm
e
d
in
it
s
knowle
dge
ba
s
e
,
whic
h
r
e
s
ul
ts
in
r
e
pe
ti
ti
ve
a
ns
we
r
s
.
T
he
s
e
c
ond
is
s
ue
is
th
a
t
J
a
c
ob
s
ometim
e
s
mi
s
unde
r
s
tood
the
c
ontext
o
f
the
que
s
t
ion
be
c
a
us
e
it
ha
s
ne
ve
r
be
e
n
lea
r
ne
d
be
f
or
e
by
th
e
W
it
.
a
i
platf
or
m.
T
h
is
is
due
to
the
pr
oblem
whe
r
e
s
im
il
a
r
s
e
ntenc
e
s
c
ould
ha
ve
the
s
a
me
mea
ning
or
c
ontext.
T
hus
,
the
W
it
.
a
i
plat
f
or
m
c
ould
mi
s
take
nly
give
dif
f
e
r
e
nt
c
ontext
or
e
ve
n
doe
s
not
r
e
c
ognize
the
c
ontext
a
t
a
ll
.
T
he
s
olut
ion
to
the
f
ir
s
t
pr
oblem
is
by
va
r
ying
th
e
a
ns
we
r
s
us
ing
a
pa
r
a
phr
a
s
e
.
I
n
na
tur
a
l
langua
ge
pr
oc
e
s
s
ing
(
NL
P
)
,
pa
r
a
phr
a
s
e
s
a
r
e
a
n
int
e
r
e
s
ti
ng
t
a
s
k
to
s
olve.
I
t
is
dif
f
icult
to
buil
d
a
pa
r
a
phr
a
s
e
r
e
c
ognit
ion
s
ys
tem
be
c
a
u
s
e
pa
r
a
phr
a
s
e
s
a
r
e
ha
r
d
to
de
f
ine
[
5
]
.
I
n
the
li
nguis
ti
c
li
ter
a
tur
e
,
pa
r
a
phr
a
s
e
s
a
r
e
s
e
ntenc
e
s
or
phr
a
s
e
s
that
ha
ve
a
n
a
ppr
oxim
a
te
e
quivale
nc
e
of
mea
ning
in
the
di
f
f
e
r
e
nt
wor
ding
[
5
]
.
T
hus
,
a
de
e
p
ne
ur
a
l
ne
twor
k
is
us
e
d
in
thi
s
s
tudy
to
ge
ne
r
a
te
the
pa
r
a
phr
a
s
e
of
the
a
ns
we
r
.
F
o
r
NL
P
tas
ks
,
r
e
c
ur
r
e
n
t
ne
ur
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
lev
e
r
e
e
:
an
ar
ti
fi
c
ial
ly
int
e
ll
igent
w
e
b
s
e
r
v
ice
for
J
ac
ob
v
oice
c
hatbot
(
Oc
tavany
)
1423
ne
twor
k
(
R
NN
)
a
r
c
hit
e
c
tur
e
g
ives
a
good
pe
r
f
o
r
manc
e
,
e
s
pe
c
ially
to
a
s
pe
c
ial
kind
o
f
R
NN
c
a
ll
e
d
L
ong
s
hor
t
-
ter
m
memor
y
(
L
S
T
M
)
,
by
r
e
duc
ing
the
pe
r
plexity
a
nd
wor
d
e
r
r
or
r
a
te
[
6
,
7
]
.
L
S
T
M
is
wid
e
ly
we
ll
known
due
to
it
s
c
a
pa
bil
it
y
o
f
lea
r
ning
long
-
ter
m
de
pe
nde
nc
ies
a
nd
r
e
duc
ing
the
va
nis
hing
gr
a
dient
pr
oblem.
L
S
T
M
c
ould
be
im
pleme
nted
to
pr
e
dict
the
ne
xt
wor
d
by
the
p
r
e
vious
wor
ds
in
langua
ge
model
ing
a
nd
ge
ne
r
a
ti
ng
text
a
nd
im
pleme
nted
in
c
ha
tbot
a
pp
li
c
a
ti
on
[
2]
.
Ya
vuz
e
t
a
l.
in
[
8
]
de
ve
lop
a
r
e
s
pons
e
ge
ne
r
a
ti
on
us
ing
L
S
T
M
a
nd
hier
a
r
c
hica
l
point
e
r
ne
twor
k
.
F
ur
ther
mor
e
,
the
r
e
is
a
n
L
S
T
M
model
that
is
us
e
d
to
ge
ne
r
a
te
pa
r
a
phr
a
s
e
by
a
dding
the
r
e
s
idue,
c
a
ll
e
d
s
tac
ke
d
r
e
s
idual
L
S
T
M
[
9
]
.
W
e
us
e
thi
s
model
be
c
a
us
e
it
h
a
s
be
tt
e
r
r
e
s
ult
s
than
S
e
que
nc
e
to
S
e
que
nc
e
,
B
i
-
dir
e
c
t
ional
L
S
T
M
,
a
tt
e
nti
on
-
ba
s
e
d
L
S
T
M
model
[
9
]
.
T
he
pr
e
-
tr
a
ined
model
of
s
tac
ke
d
r
e
s
idual
L
S
T
M
f
r
om
[
9
]
is
im
ple
mente
d
to
ge
ne
r
a
te
pa
r
a
phr
a
s
e
of
a
ns
we
r
s
in
thi
s
p
a
pe
r
.
I
n
the
f
a
c
e
of
the
une
xpe
c
te
d
r
e
s
ult
s
of
int
e
nt
a
nd
e
nti
ti
e
s
,
it
is
ne
c
e
s
s
a
r
y
to
upda
te
the
knowle
dge
in
the
W
it
.
a
i
platf
or
m
f
r
om
the
his
tor
y
of
que
s
ti
o
ns
.
B
a
s
e
d
on
the
s
tudy
a
bout
J
a
c
ob,
ther
e
a
r
e
c
on
ve
r
s
a
ti
on
logs
that
r
e
c
or
d
a
ll
the
c
onve
r
s
a
ti
o
ns
be
twe
e
n
J
a
c
ob
a
nd
the
us
e
r
.
S
o,
we
pr
opos
e
the
s
olut
ion
to
pr
e
ve
nt
a
n
a
dmi
nis
tr
a
tor
to
manua
ll
y
r
e
a
d
a
ll
of
the
que
s
ti
ons
in
the
c
onve
r
s
a
ti
on
logs
b
y
e
xtr
a
c
ti
ng
the
s
um
mar
y
of
que
s
ti
ons
,
whic
h
s
e
lec
ts
the
mos
t
f
r
e
que
ntl
y
a
s
ke
d
que
s
ti
ons
f
r
om
the
e
nti
r
e
c
onve
r
s
a
ti
on
logs
[
10
]
.
I
de
a
ll
y
,
the
e
xtr
a
c
ti
ve
su
mm
a
r
iza
ti
on
s
hould
c
ontain
a
r
oun
d
20%
o
f
the
s
e
nt
e
nc
e
s
f
r
om
the
e
nti
r
e
text
[
1
1
]
.
T
he
r
e
a
r
e
s
ix
a
lgor
it
hms
f
or
e
xtr
a
c
ti
ve
s
umm
a
r
iza
ti
on
that
h
a
ve
be
e
n
e
va
luate
d
by
Vic
tor
e
t
a
l
in
[
1
2
]
,
whic
h
a
r
e
L
uhn,
T
e
xtR
a
nk,
L
e
xR
a
nk,
L
S
A,
S
umB
a
s
ic,
a
nd
KL
S
u
m.
B
a
s
e
d
on
the
r
e
s
ult
s
,
L
uhn
a
nd
T
e
xtR
a
nk
ha
v
e
the
be
s
t
pe
r
f
or
manc
e
to
ge
t
the
e
xt
r
a
c
ti
ve
s
umm
a
r
y
f
o
r
th
e
s
pe
e
c
h
-
to
-
text
c
a
s
e
[
1
2
]
.
B
e
twe
e
n
thos
e
both
a
lgor
it
hms
,
we
c
hoos
e
T
e
xtR
a
nk
be
c
a
us
e
the
T
e
xtR
a
nk
a
lg
or
it
hm
is
a
gr
a
ph
-
ba
s
e
d
r
a
nking
a
lgor
it
hm
that
ha
s
be
e
n
pr
ove
n
to
be
s
uc
c
e
s
s
f
ul
f
or
the
identif
ica
ti
on
of
th
e
mos
t
im
po
r
tant
o
r
r
e
leva
nt
s
e
ntenc
e
s
(
ve
r
tex)
in
the
text
(
gr
a
ph)
[
1
3
]
.
Ac
c
or
ding
to
[
10
,
1
4
]
T
e
xtR
a
nk
a
lgor
it
hm
with
c
os
ine
s
im
il
a
r
it
y
us
ing
W
or
d2V
e
c
c
ould
e
nha
nc
e
the
r
a
nking
pr
oc
e
s
s
.
2.
RE
S
E
AR
CH
M
E
T
HO
D
T
his
s
e
c
ti
on
c
ontains
a
br
ief
e
xplana
ti
on
o
f
L
S
T
M
a
s
the
main
method
im
pleme
nted
in
thi
s
s
tudy.
T
he
n,
we
de
s
c
r
ibe
T
e
xtR
a
nk
a
lgo
r
it
hm,
C
os
ine
S
im
il
a
r
it
y,
a
nd
W
o
r
d
E
mbedding.
I
n
or
de
r
to
that,
we
a
ls
o
e
xplain
our
p
r
opos
e
d
method.
2.
1.
L
on
g
s
h
or
t
-
t
e
r
m
m
e
m
or
y
(
L
S
T
M
)
L
S
T
M
is
a
va
r
iant
of
the
r
e
c
ur
r
e
nt
ne
twor
k
whic
h
is
di
f
f
e
r
e
nt
f
r
om
the
f
e
e
d
-
f
or
wa
r
d
ne
twor
k
a
nd
is
int
r
oduc
e
d
by
[
1
5
]
.
R
e
c
ur
r
e
nt
ne
twor
k
f
e
e
ds
it
s
output
s
ba
c
k
int
o
it
s
own
input
s
s
o
the
r
e
s
p
ons
e
of
the
ne
twor
k
to
a
given
input
may
de
pe
nd
on
p
r
e
vious
input
s
[
1
6
]
.
L
T
S
M
c
omput
e
s
the
hidden
s
t
a
te
ℎ
by
a
dding
a
memor
y
c
e
ll
a
t
e
ve
r
y
ti
me
s
tep
[
9
]
.
W
h
e
n
the
unit
c
o
mput
e
s
the
memor
y
c
e
ll
a
nd
hidden
s
tate
a
t
ti
me
s
tep
t
,
i
t
c
ons
ider
s
the
inpu
t
s
tate
a
t
ti
me
s
tep
,
the
hidden
s
tate
ℎ
−
1
,
a
nd
the
memor
y
c
e
ll
−
1
a
t
ti
me
s
tep
[
9
]
.
L
S
T
M
ha
s
th
r
e
e
ga
tes
that
e
a
c
h
ga
te
c
ons
is
ts
of
one
s
igm
oid
laye
r
a
nd
po
int
w
is
e
mul
ti
pli
c
a
ti
on
ope
r
a
ti
on
whic
h
known
a
s
f
or
ge
t
ga
te
(
)
,
input
ga
te
(
)
,
a
nd
output
ga
te
(
)
[
1
7
]
.
T
he
ga
t
e
s
a
r
e
de
s
c
r
ibed
in
F
igu
r
e
1.
F
igur
e
1.
L
S
T
M
c
e
ll
c
ontains
f
or
ge
t
ga
te
,
input
ga
t
e
,
ne
w
memor
y
c
e
ll
,
a
nd
output
ga
te
[
1
7
]
2.
1.
1.
S
t
ac
k
e
d
r
e
s
id
u
al
L
S
T
M
S
tac
ke
d
R
e
s
idual
L
S
T
M
is
a
model
that
wa
s
pr
op
os
e
d
by
P
r
a
ka
s
h
e
t
a
l.
a
nd
the
a
ddit
ion
of
r
e
s
idual
c
onne
c
ti
ons
to
ge
ne
r
a
te
pa
r
a
phr
a
s
e
a
nd
thi
s
mode
l
c
a
n
he
lp
o
ve
r
c
ome
a
de
gr
a
da
ti
on
pr
ob
lem
[
9
]
.
R
e
s
idual
c
onne
c
ti
ons
a
r
e
a
dde
d
a
f
ter
e
ve
r
y
n
lay
e
r
s
a
s
the
point
wis
e
a
ddit
ion
[
9
]
.
F
igur
e
2
s
hows
that
the
r
e
s
idual
c
onne
c
ti
ons
a
r
e
a
dde
d
a
f
ter
e
ve
r
y
two
laye
r
s
a
s
the
point
wis
e
a
ddit
ion.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
422
-
1
432
1424
F
igur
e
2.
A
uni
t
of
S
tac
ke
d
R
e
s
idual
L
S
T
M
[
9
]
2.
2.
T
e
xt
Rank
T
e
xtR
a
nk
is
a
gr
a
ph
-
ba
s
e
d
r
a
nking
a
lgor
it
hm
f
or
t
he
identif
ica
ti
on
of
the
mos
t
im
por
tant
or
r
e
leva
nt
s
e
ntenc
e
s
in
the
text
[
1
3
]
.
One
ve
r
tex
r
e
pr
e
s
e
nts
a
s
e
ntenc
e
a
nd
the
e
dge
s
r
e
pr
e
s
e
nt
the
r
e
lation
s
c
or
e
(
we
ight
)
be
twe
e
n
two
s
e
ntenc
e
s
.
A
ve
r
tex
wi
th
the
h
ighes
t
s
c
or
e
of
T
e
xtR
a
nk
is
the
mos
t
r
e
l
e
va
nt
or
im
por
tant
in
the
gr
a
ph
.
T
e
xtR
a
nk
is
a
we
ight
e
d
dir
e
c
ted
gr
a
ph
=
(
,
)
,
whe
r
e
the
gr
a
ph
c
ons
is
ts
of
a
s
e
t
of
ve
r
ti
c
e
s
a
nd
a
s
e
t
o
f
e
dge
s
.
F
or
a
given
ve
r
tex
,
(
)
r
e
pr
e
s
e
nts
the
s
e
t
of
ve
r
ti
c
e
s
that
point
to
ve
r
tex
,
a
nd
(
)
r
e
p
r
e
s
e
nts
the
s
e
t
of
ve
r
ti
c
e
s
that
p
oint
f
r
om
ve
r
tex
.
T
he
we
ight
f
r
om
ve
r
tex
i
to
ve
r
tex
j
is
r
e
pr
e
s
e
nted
a
s
.
T
he
f
o
r
mul
a
o
f
T
e
xtR
a
nk
c
a
n
be
de
f
ined
a
s
s
hown
in
(
1
)
[
1
8
]
,
(
)
=
(
1
−
)
+
∑
∈
(
)
∑
∈
(
)
(
)
(
1)
in
(
1)
,
(
)
r
e
pr
e
s
e
nts
the
s
c
or
e
of
ve
r
tex
;
is
the
da
mpi
ng
f
a
c
tor
with
va
lue
is
be
twe
e
n
0
to
1.
Nor
mally,
the
d
a
mpi
ng
f
a
c
tor
is
s
e
t
to
0.
85
[
1
3
].
2.
2.
1.
Wor
d
e
m
b
e
d
d
in
g
W
or
d
e
mbedding
is
a
ve
c
tor
r
e
pr
e
s
e
ntation
o
f
wor
ds
ba
s
e
d
on
the
c
ontext
of
the
s
e
ntenc
e
s
or
s
e
mantic
r
e
lations
hips
be
twe
e
n
wor
ds
.
T
he
ve
c
to
r
c
ontains
r
e
a
l
numbe
r
[
1
9
]
.
M
ikol
ov
e
t
a
l.
pr
op
os
e
d
two
models
that
f
oc
us
on
lea
r
ning
wor
d
ve
c
tor
s
whi
c
h
a
r
e
c
onti
nuous
ba
g
-
of
-
wor
ds
(
C
B
oW
)
a
nd
C
o
nti
nuous
S
kip
-
gr
a
m
(
S
G)
a
s
s
hown
in
F
igur
e
3.
T
he
s
e
two
models
c
a
ll
e
d
W
or
d2Ve
c
.
C
B
oW
is
opti
mi
z
e
d
t
o
pr
e
dict
a
wor
d
ba
s
e
d
on
wor
ds
a
r
ound
it
or
the
c
ontext
.
S
G
is
opti
mi
z
e
d
to
p
r
e
dict
the
c
ontext
ba
s
e
d
on
the
c
ur
r
e
nt
wor
d
[
20
].
F
igur
e
3.
C
onti
nuous
ba
g
-
of
-
wor
ds
(
lef
t)
a
nd
s
kip
-
gr
a
m
(
r
ight
)
[
20
]
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
lev
e
r
e
e
:
an
ar
ti
fi
c
ial
ly
int
e
ll
igent
w
e
b
s
e
r
v
ice
for
J
ac
ob
v
oice
c
hatbot
(
Oc
tavany
)
1425
2.
2.
2.
Cos
in
e
s
im
il
ar
i
t
y
C
os
ine
S
im
il
a
r
it
y
c
a
lcula
tes
the
s
im
il
a
r
it
y
by
mea
s
ur
ing
the
c
os
ine
of
th
e
a
ngle
be
twe
e
n
tw
o
ve
c
tor
s
.
A
r
e
s
ult
is
a
number
be
twe
e
n
z
e
r
o
a
nd
one
.
I
f
the
r
e
s
ult
is
c
los
e
to
one
,
the
mo
r
e
s
im
il
a
r
it
s
two
ve
c
tor
s
[
2
1
]
.
I
n
thi
s
pa
pe
r
,
the
ve
c
to
r
s
r
e
pr
e
s
e
nt
t
he
s
e
ntenc
e
ve
c
tor
s
whic
h
a
r
e
c
a
lcula
ted
us
ing
W
or
d2Ve
c
.
T
he
s
im
il
a
r
it
y
be
tw
e
e
n
ve
c
tor
i
⃗
a
nd
ve
c
tor
j
⃗
c
a
n
be
de
f
ined
a
s
s
hown
in
(
2)
:
(
,
)
=
(
⃗
,
⃗
)
=
⃗
.
⃗
|
|
⃗
|
|
2
∗
|
|
⃗
|
|
2
(
2)
in
F
or
mul
a
2,
s
ymbol
.
a
t
⃗
.
⃗
de
notes
the
dot
-
pr
oduc
t
of
ve
c
tor
⃗
a
nd
ve
c
tor
⃗
.
T
he
notation
of
|
|
⃗
|
|
2
r
e
pr
e
s
e
nts
the
ve
c
tor
magnitude
o
f
ve
c
to
r
⃗
.
I
n
[
2
2
]
,
C
os
ine
S
im
il
a
r
it
y
is
a
ls
o
us
e
d
to
mea
s
ur
e
the
s
e
ntenc
e
s
i
mi
lar
it
y
with
the
M
a
laya
lam
langua
ge
.
2.
3.
P
r
op
os
e
d
m
e
t
h
od
2.
3.
1.
De
s
ign
T
he
de
s
ign
of
C
leve
r
e
e
take
s
int
o
c
ons
ider
a
ti
on
s
e
ve
r
a
l
J
a
c
ob
s
pe
c
if
ica
ti
ons
.
J
a
c
ob
us
e
s
E
ngli
s
h
a
s
the
langua
ge
a
nd
W
it
.
a
i
platf
o
r
m
to
ge
t
the
int
e
nt
a
nd
e
nti
ti
e
s
f
r
om
the
c
onve
r
s
a
ti
on.
J
a
c
ob’
s
knowle
dge
ba
s
e
is
s
tor
e
d
us
ing
the
M
yS
QL
da
taba
s
e
a
nd
ther
e
is
no
im
pleme
ntation
of
a
ny
a
r
ti
f
icia
l
int
e
ll
igenc
e
f
e
a
tur
e
s
in
J
a
c
ob
e
xc
e
pt
f
or
the
W
it
.
a
i
platf
o
r
m.
J
a
c
ob
i
s
a
we
b
-
ba
s
e
d
a
ppli
c
a
ti
on
buil
t
us
ing
P
HP
pr
og
r
a
mm
ing
langua
ge
a
nd
L
a
r
a
ve
l
F
r
a
mew
or
k.
J
a
c
ob
c
ould
on
ly
r
e
s
pond
to
que
s
ti
ons
that
ha
ve
a
ns
we
r
s
in
the
d
a
taba
s
e
,
f
ur
ther
mor
e
,
one
que
s
ti
on
ha
s
only
one
a
ns
we
r
.
T
he
W
it
.
a
i
platf
o
r
m
s
ometim
e
s
c
ould
give
d
if
f
e
r
e
nt
c
ontext
or
doe
s
not
r
e
c
ognize
the
c
ontext
f
or
s
im
il
a
r
qu
e
s
ti
ons
.
T
he
r
e
f
or
e
,
two
a
r
ti
f
icia
l
int
e
ll
igenc
e
f
e
a
tur
e
s
a
r
e
pr
opos
e
d
to
be
a
dde
d
to
J
a
c
ob.
T
he
pur
pos
e
is
t
o
make
J
a
c
ob
mo
r
e
int
e
l
li
ge
nt
by
r
e
plyi
ng
to
the
s
a
me
or
s
im
il
a
r
que
s
ti
ons
with
a
va
r
iation
of
a
ns
we
r
s
.
I
n
a
ddit
ion
to
that
is
to
upda
te
the
knowle
dge
b
a
s
e
int
o
the
W
it
.
a
i
plat
f
or
m
.
T
he
C
leve
r
e
e
is
de
s
igned
a
s
a
we
b
s
e
r
vice
with
P
ython
pr
ogr
a
mm
ing
langua
ge
a
nd
F
las
k
f
r
a
mew
or
k
to
a
ll
ow
the
a
ppli
c
a
ti
on
to
be
platf
o
r
m
a
nd
tec
hnology
indepe
nde
nt.
T
he
we
b
s
e
r
vice
de
s
igned
with
f
ou
r
a
c
c
e
s
s
ibl
e
UR
L
s
is
s
hown
in
F
igur
e
4.
T
he
UR
L
s
a
r
e
f
or
pa
r
a
phr
a
s
e
ge
ne
r
a
ti
on,
que
s
ti
ons
s
umm
a
r
iza
ti
on,
a
dd
tr
a
ini
ng
da
ta
,
a
nd
t
r
a
ini
ng
the
model
.
W
he
n
the
da
ta
is
s
e
nt
to
the
we
b
s
e
r
vice
,
the
we
b
s
e
r
vice
is
de
s
igned
to
r
e
c
e
ive
the
da
ta
us
ing
P
O
S
T
method
r
e
que
s
t.
T
hus
,
t
h
is
we
b
s
e
r
vice
c
ould
be
a
c
c
e
s
s
e
d
by
a
ny
we
b
a
ppli
c
a
ti
on
not
only
J
a
c
ob
.
Additi
ona
l
c
ha
nge
s
a
r
e
made
in
J
a
c
ob’
s
knowle
dge
b
a
s
e
,
it
is
to
a
dd
thr
e
e
table
s
to
s
tor
e
the
r
e
s
ult
s
of
pa
r
a
phr
a
s
e
ge
ne
r
a
ti
on
a
nd
s
umm
a
r
iza
ti
on.
F
igur
e
4.
C
leve
r
e
e
we
b
s
e
r
vice
model
C
leve
r
e
e
us
e
s
a
pr
e
-
tr
a
ined
model
S
tac
ke
d
R
e
s
idual
L
S
T
M
f
o
r
pa
r
a
ph
r
a
s
e
ge
ne
r
a
ti
on.
I
n
F
igur
e
4,
ne
w
tr
a
ini
ng
da
ta
c
ould
be
a
dde
d
via
UR
L
“
/add
ingdata
”
.
T
his
a
c
ti
on
c
ould
only
be
done
us
ing
a
n
a
dmi
n
r
ole.
T
he
a
dmi
n
c
ould
c
hoos
e
to
tr
a
in
the
mod
e
l
with
the
ne
w
da
tas
e
t
to
incr
e
a
s
e
the
knowle
dge
a
nd
the
va
r
iati
on
o
f
s
e
ntenc
e
s
.
T
his
pr
oc
e
s
s
r
uns
in
the
ba
c
kgr
ound
s
o
a
dmi
n
c
a
n
do
other
a
c
ti
vit
ies
.
W
he
n
J
a
c
ob
c
a
ll
s
the
UR
L
“
/t
r
a
ini
ng”
a
nd
the
tr
a
ini
ng
pr
oc
e
s
s
is
r
unn
ing,
we
c
r
e
a
te
a
ne
w
voc
a
bular
y
ba
s
e
d
on
the
ne
w
da
tas
e
t.
Af
te
r
that
,
the
we
ight
s
in
ne
ur
a
l
ne
twor
ks
a
r
e
a
lwa
ys
r
e
-
ini
ti
a
li
z
e
d.
All
of
the
c
ha
r
a
c
ter
s
a
r
e
in
lowe
r
c
a
s
e
a
nd
punc
tuation
is
not
take
n
int
o
a
c
c
ount
f
or
the
input
a
nd
output
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
422
-
1
432
1426
the
model.
I
f
the
wo
r
d
is
not
f
ound
in
the
voc
a
bular
y
da
taba
s
e
,
the
wor
d
is
s
e
t
to
<
UN
K>
to
r
e
pr
e
s
e
nt
a
n
unknown
wor
d
a
s
in
[
9
]
.
T
his
a
im
s
to
lea
r
n
ne
w
wor
ds
a
nd
ne
w
s
e
ntenc
e
s
.
T
he
t
r
a
ini
ng
pr
oc
e
s
s
r
uns
in
the
ba
c
kgr
ound
s
o
a
dmi
n
c
a
n
do
other
a
c
ti
vit
ies
.
J
a
c
ob
r
e
c
e
ives
the
s
e
nten
c
e
or
que
s
ti
on
f
r
om
the
us
e
r
,
a
nd
s
e
nds
a
r
e
que
s
t
to
UR
L
“
/par
a
phr
a
s
e
”
.
T
he
n,
s
tar
ti
ng
with
lowe
r
c
a
s
e
a
ll
of
the
wor
ds
in
t
he
s
e
ntenc
e
.
Ne
xt,
load
the
pa
r
a
phr
a
s
e
model
that
ha
s
be
e
n
tr
a
ined
be
f
or
e
.
T
he
model
gives
the
pr
e
diction
a
s
a
r
e
s
ult
of
pa
r
a
phr
a
s
e
ba
s
e
d
on
the
s
e
nt
e
nc
e
a
s
s
hown
in
F
igur
e
5
.
T
he
pr
e
diction
is
a
li
s
t
o
f
nu
mber
s
.
T
he
number
r
e
pr
e
s
e
nts
the
id
o
f
wo
r
d
s
o
e
ve
r
y
wor
d
ha
s
a
dif
f
e
r
e
nt
nu
mber
r
e
pr
e
s
e
ntation.
T
he
n
,
the
num
be
r
is
c
onve
r
ted
to
wo
r
d
a
nd
c
onc
a
tena
te
a
ll
o
f
th
e
wor
ds
.
L
a
s
t,
the
pa
r
a
phr
a
s
e
d
s
e
ntenc
e
is
s
tor
e
d
in
J
a
c
ob’
s
da
taba
s
e
s
o
the
a
dmi
n
c
a
n
va
li
da
te
a
nd
de
lete
the
r
e
s
ult
in
the
J
a
c
ob
a
dmi
n
s
ys
tem.
I
f
ther
e
a
r
e
e
r
r
or
s
a
n
d
a
dmi
n
s
ti
ll
unde
r
s
tand
the
c
ontext,
a
dmi
n
c
a
n
c
or
r
e
c
t
the
s
e
ntenc
e
then
va
li
da
tes
it
s
o
the
s
e
ntenc
e
c
a
n
be
us
e
d
a
s
a
n
a
ns
we
r
.
T
he
e
r
r
or
s
of
s
e
ntenc
e
s
c
a
n
be
gr
a
mm
a
ti
c
a
l
e
r
r
or
s
,
s
yntac
t
ica
l
e
r
r
o
r
s
,
a
nd
s
e
manti
c
e
r
r
or
s
.
F
igur
e
5.
F
lowc
ha
r
t
of
pa
r
a
phr
a
s
e
ge
ne
r
a
ti
on
T
he
r
e
a
r
e
s
e
ve
r
a
l
s
teps
to
do
que
s
ti
on
s
umm
a
r
iza
t
ion
whe
n
J
a
c
ob
c
a
ll
s
the
U
R
L
“
/
s
umm
a
r
y”
a
nd
it
de
s
c
r
ibes
in
F
igur
e
6.
F
i
r
s
t,
ge
t
the
c
onve
r
s
a
ti
on
l
og
f
il
e
s
that
ha
ve
n’
t
be
e
n
s
umm
a
r
ize
d.
T
his
c
onve
r
s
a
ti
on
log
f
i
les
a
r
e
s
tor
e
d
whe
n
ther
e
is
a
us
e
r
int
e
r
a
c
ts
with
J
a
c
ob.
S
e
c
ond,
do
the
s
umm
a
r
y
pr
e
pr
oc
e
s
s
ing.
I
n
s
umm
a
r
y
pr
e
pr
oc
e
s
s
ing,
we
r
e
move
a
ll
the
pun
c
tuations
a
nd
number
s
,
tokeniz
e
the
s
e
ntenc
e
s
int
o
wor
ds
,
c
ha
nge
a
ll
o
f
the
wo
r
ds
to
lowe
r
c
a
s
e
,
r
e
mo
ve
s
topwor
ds
,
a
nd
load
the
pr
e
-
tr
a
ined
wor
d
ve
c
tor
s
.
T
he
n,
c
a
lcula
te
the
s
e
ntenc
e
ve
c
tor
s
by
a
ve
r
a
ging
the
tot
a
l
of
wor
d
ve
c
tor
s
f
or
e
a
c
h
s
e
ntenc
e
.
T
hi
r
d,
c
r
e
a
te
the
s
im
il
a
r
it
y
matr
ix
with
s
ize
n
x
n
,
whe
r
e
n
is
th
e
tot
a
l
o
f
s
e
ntenc
e
s
.
T
he
va
lue
f
or
e
a
c
h
r
ow
a
nd
c
olum
n
is
c
a
lcula
ted
us
ing
the
C
os
ine
S
im
il
a
r
i
ty
method
th
a
t
r
e
pr
e
s
e
nts
the
s
im
il
a
r
it
y
of
e
a
c
h
two
-
s
e
ntenc
e
ve
c
tor
s
.
T
he
r
ow
in
the
matr
ix
r
e
pr
e
s
e
nts
the
f
ir
s
t
s
e
ntenc
e
a
nd
the
c
olu
mn
r
e
p
r
e
s
e
nts
the
s
e
c
ond
s
e
ntenc
e
that
wa
nts
to
c
ompar
e
.
F
our
th
,
do
s
umm
a
r
y
e
xtr
a
c
ti
on
us
in
g
the
T
e
xtR
a
nk
a
lgor
it
hm.
T
he
we
ight
s
in
e
ve
r
y
e
dge
of
the
gr
a
ph
us
e
the
va
lue
f
r
om
the
s
im
il
a
r
it
y
matr
ix
.
T
he
n,
c
a
lcula
te
the
s
c
or
e
s
f
or
e
ve
r
y
e
dge
us
ing
(
1
)
unti
l
it
r
e
a
c
he
s
the
c
onve
r
ge
nc
e
[
1
3
]
.
T
he
las
t,
c
hoos
e
25%
s
e
ntenc
e
s
whic
h
ha
ve
the
highes
t
s
c
o
r
e
f
r
om
the
e
nti
r
e
que
s
ti
ons
or
r
e
que
s
ts
that
a
s
k
by
us
e
r
[
10
,
1
1
]
.
S
a
me
a
s
the
r
e
s
ult
o
f
pa
r
a
phr
a
s
e
,
the
r
e
s
ult
s
of
que
s
ti
on
s
umm
a
r
iza
ti
on
a
r
e
s
tor
e
d
in
the
da
ta
ba
s
e
a
nd
a
dmi
n
c
a
n
va
li
da
te
or
de
lete
th
e
r
e
s
ult
of
que
s
ti
on
s
umm
a
r
iza
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
lev
e
r
e
e
:
an
ar
ti
fi
c
ial
ly
int
e
ll
igent
w
e
b
s
e
r
v
ice
for
J
ac
ob
v
oice
c
hatbot
(
Oc
tavany
)
1427
F
igur
e
6.
F
lowc
ha
r
t
of
s
umm
a
r
iza
ti
on
W
e
us
e
the
W
hit
e
box
tes
ti
ng
a
ppr
oa
c
h
to
mea
s
ur
e
the
im
pleme
ntation
of
the
C
os
ine
S
im
il
a
r
it
y
method.
I
t
is
be
c
a
us
e
thi
s
tes
ti
ng
is
s
uit
a
ble
to
tes
t
the
a
lgor
it
h
m.
Unlike
the
othe
r
r
e
s
e
a
r
c
h
a
bout
Ar
ti
f
icia
l
I
ntelli
ge
nc
e
ge
ne
r
a
ll
y,
in
thi
s
pa
pe
r
,
we
us
e
the
T
e
c
hnology
Ac
c
e
ptanc
e
M
ode
l
(
T
AM
)
a
s
a
n
e
va
luation
method.
T
AM
is
us
e
d
to
pr
e
dict
the
a
c
c
e
ptanc
e
of
tec
hnology
in
a
n
or
ga
niza
ti
on
.
T
he
r
e
s
ult
s
o
f
T
AM
a
r
e
de
ter
mi
ne
d
ba
s
e
d
on
two
pe
r
c
e
ived
va
r
iable
s
whic
h
a
r
e
P
e
r
c
e
ived
Us
e
f
ulne
s
s
a
nd
P
e
r
c
e
ived
E
a
s
e
o
f
Us
e
[
2
3
]
.
W
e
us
e
the
ini
ti
a
l
s
c
a
le
it
e
ms
f
or
P
e
r
c
e
iv
e
d
Us
e
f
ulnes
s
a
nd
P
e
r
c
e
ived
E
a
s
e
of
Us
e
[
2
4
,
25
].
T
r
a
ini
ng
da
ta
f
or
the
S
tac
ke
d
R
e
s
idual
L
S
T
M
mod
e
l
is
s
tor
e
d
in
T
e
xt
Doc
ument
s
,
c
ons
is
ti
ng
of
tr
a
in
s
our
c
e
a
nd
tr
a
in
tar
ge
t.
T
he
r
e
is
no
tes
t
da
tas
e
t
f
or
thi
s
model,
but
f
or
the
e
va
luation,
we
us
e
a
n
i
nter
view
with
J
a
c
ob
a
dmi
ns
to
know
the
pa
r
a
phr
a
s
e
r
e
s
ult
s
a
r
e
f
e
a
s
ibl
e
or
not
.
Othe
r
than
pa
r
a
phr
a
s
e
r
e
s
ult
s
,
we
do
a
n
int
e
r
view
to
s
uppor
t
the
e
va
luation
r
e
s
ult
s
us
ing
T
AM
a
nd
to
know
the
s
umm
a
r
iza
ti
on
r
e
s
ult
s
a
r
e
f
e
a
s
ibl
e
or
not.
E
va
luations
us
e
T
AM
a
nd
int
e
r
view
,
we
do
thes
e
two
e
va
luations
with
th
r
e
e
a
dmi
ns
.
2.
3.
2.
I
m
p
lem
e
n
t
at
ion
T
his
pa
r
t
e
xplains
the
int
e
gr
a
ti
on
o
f
C
leve
r
e
e
modul
e
int
o
J
a
c
ob
a
ppli
c
a
ti
on,
the
r
e
s
ult
of
the
im
pleme
ntation
of
S
tac
ke
d
R
e
s
idual
L
S
T
M
,
C
os
ine
S
im
il
a
r
it
y,
a
nd
T
e
xtR
a
nk.
C
leve
r
e
e
modul
e
is
buil
t
li
ke
the
de
s
c
r
ipt
ion
in
the
r
e
s
e
a
r
c
h
method
whic
h
i
s
buil
t
a
s
a
we
b
s
e
r
vice
a
nd
the
r
e
s
ult
is
s
hown
in
F
igur
e
7.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
422
-
1
432
1428
F
igur
e
7.
S
c
r
e
e
ns
hot
of
c
omm
a
nd
p
r
ompt
that
s
hows
the
C
leve
r
e
e
s
e
r
ve
r
is
on
2.
3.
3.
I
n
t
e
gr
at
ion
in
j
ac
ob
a
d
m
in
s
ys
t
e
m
T
he
int
e
gr
a
t
ion
in
th
is
pa
r
t
mea
ns
the
ti
me
whe
n
C
leve
r
e
e
is
c
a
ll
e
d
by
J
a
c
ob.
J
a
c
ob
c
a
ll
s
the
UR
L
“
/par
a
phr
a
s
e
”
whe
n
the
a
dmi
n
wa
nts
t
o
c
r
e
a
te
a
ne
w
a
ns
we
r
a
nd
whe
n
J
a
c
ob
gives
the
a
n
s
we
r
to
the
us
e
r
that
is
s
hown
in
F
igur
e
8.
T
he
n
,
J
a
c
ob
c
a
ll
s
the
UR
L
“
/s
umm
a
r
y”
whe
n
a
dmi
n
ope
ns
the
logi
n
pa
ge
whic
h
is
s
hown
in
F
igu
r
e
9
.
J
a
c
ob
c
a
ll
s
the
UR
L
“
/addingd
a
ta”
whe
n
a
dmi
n
c
r
e
a
tes
a
ne
w
a
ns
we
r
s
o
the
s
e
ntenc
e
s
a
r
e
a
dde
d
to
t
r
a
in
s
our
c
e
a
nd
tr
a
in
tar
ge
t
da
ta.
B
a
s
e
d
on
the
a
ns
we
r
that
c
r
e
a
ted
by
a
dmi
n
,
C
leve
r
e
e
woul
d
c
r
e
a
te
the
pa
r
a
phr
a
s
e
d
s
e
ntenc
e
.
T
he
n,
a
dmi
ns
c
a
n
a
c
c
e
pt
or
r
e
vis
e
the
s
e
ntenc
e
.
L
a
s
t,
J
a
c
ob
c
a
ll
s
t
he
UR
L
“
/t
r
a
ini
ng”
whe
n
a
n
a
dmi
n
wa
nts
to
upda
te
the
kno
wle
dge
a
nd
va
r
iation
o
f
a
ns
we
r
s
ba
s
e
d
on
a
ne
w
d
a
tas
e
t.
F
igur
e
8.
A
c
a
s
e
whe
n
UR
L
“
/par
a
phr
a
s
e
”
is
c
a
ll
e
d
F
igur
e
9.
T
he
c
a
s
e
whe
n
UR
L
“
/s
umm
a
r
y”
is
c
a
ll
e
d
2.
3.
4.
I
m
p
lem
e
n
t
at
ion
of
p
r
e
-
t
r
ain
e
d
m
od
e
l
of
s
t
ac
k
e
d
r
e
s
id
u
al
L
S
T
M
T
he
im
pleme
ntation
o
f
thi
s
pr
e
-
tr
a
ined
model
is
u
s
ing
T
e
ns
or
f
low
pa
c
ka
ge
in
P
ython
p
r
ogr
a
mm
ing
langua
ge
.
T
he
input
a
nd
output
dim
e
ns
ions
a
r
e
th
e
s
a
m
e
,
the
dim
e
ns
ion
s
ize
is
256
.
T
he
dim
e
ns
ion
of
wor
d
e
mbedding
is
the
s
a
me
a
s
the
input
a
nd
out
put
d
im
e
ns
ions
.
I
n
tr
a
ini
ng,
the
model
us
e
s
0.
001
wit
h
a
f
ixed
va
lue
a
s
the
lea
r
ning
r
a
te
a
nd
Ada
m
opt
im
ize
r
.
T
h
e
ba
tch
s
ize
is
s
e
t
to
32
a
nd
the
number
of
it
e
r
a
ti
o
ns
is
s
e
t
to
1
,
500.
E
ve
r
y
two
L
S
T
M
la
ye
r
is
a
dde
d
with
th
e
r
e
s
idue
a
s
in
[
9
]
.
T
he
tot
a
l
tr
a
i
ning
da
ta
is
516
s
e
ntenc
e
s
a
nd
the
tot
a
l
voc
a
bular
y
is
901
wo
r
ds
,
but
it
c
a
n
incr
e
a
s
e
if
a
dmi
n
a
dds
the
ne
w
s
e
ntenc
e
s
of
the
a
ns
we
r
.
T
he
tr
a
ini
ng
da
ta
c
ontains
s
e
ntenc
e
s
a
bout
th
e
Dua
l
De
gr
e
e
pr
ogr
a
m
of
I
nf
or
matics
in
Un
iver
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
a
nd
it
s
ge
t
f
r
om
int
e
r
view
with
the
M
a
r
ke
ti
ng
Divis
ion
of
Unive
r
s
it
a
s
M
u
lt
im
e
dia
Nus
a
ntar
a
.
B
e
f
or
e
pr
oc
e
s
s
the
s
e
ntenc
e
int
o
the
model,
we
ne
e
d
to
do
pr
e
pr
oc
e
s
s
ing.
I
n
the
pr
e
pr
oc
e
s
s
ing,
we
us
e
nlt
k
pa
c
ka
ge
to
r
e
move
punc
tuation
a
n
d
tokeniz
e
the
s
e
ntenc
e
to
w
or
ds
.
F
igur
e
10
r
e
pr
e
s
e
nts
the
indi
c
a
tor
whe
n
th
e
tr
a
ini
ng
pr
oc
e
s
s
is
r
unnin
g.
I
n
the
tr
a
ini
ng
pr
oc
e
s
s
,
we
s
how
the
s
our
c
e
,
ta
r
ge
t,
a
nd
pr
e
dict
s
e
ntenc
e
.
T
his
p
r
oc
e
s
s
is
r
unning
in
the
ba
c
kgr
ound
f
o
r
the
a
dmi
n
is
tr
a
tor
to
c
a
r
r
y
out
o
ther
a
c
t
ivi
ti
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
lev
e
r
e
e
:
an
ar
ti
fi
c
ial
ly
int
e
ll
igent
w
e
b
s
e
r
v
ice
for
J
ac
ob
v
oice
c
hatbot
(
Oc
tavany
)
1429
F
igur
e
10.
T
r
a
ini
ng
p
r
oc
e
s
s
of
s
tac
ke
d
r
e
s
idual
L
S
T
M
2.
3.
5.
I
m
p
lem
e
n
t
at
ion
of
c
os
in
e
s
im
il
ar
it
y
an
d
T
e
xt
Rank
T
he
im
pleme
ntation
of
the
C
os
ine
S
im
il
a
r
it
y
me
thod
us
e
s
numpy
pa
c
ka
ge
f
or
pr
e
pr
oc
e
s
s
ing
a
nd
c
a
lcula
ti
on
ve
c
tor
s
of
e
a
c
h
s
e
ntenc
e
pa
ir
s
.
T
e
xt
R
a
nk
a
lgor
it
hm
is
im
pleme
nted
us
ing
ne
two
r
kx
pa
c
ka
ge
.
I
n
pr
e
pr
oc
e
s
s
ing
da
ta
of
s
umm
a
r
iza
ti
on
f
e
a
tur
e
is
us
e
d
nlt
k
pa
c
ka
ge
to
r
e
move
s
topwor
ds
a
nd
tokeniz
e
s
e
ntenc
e
to
li
s
t
of
w
or
ds
.
3.
RE
S
UL
T
S
AN
D
AN
AL
YSI
S
3.
1.
T
e
s
t
in
g
3.
1.
1.
T
e
s
t
in
g
f
or
im
p
lem
e
n
t
at
ion
o
f
c
os
in
e
s
im
il
ar
it
y
B
a
s
e
d
on
the
s
c
e
na
r
io
in
T
a
ble
1,
we
ha
ve
two
v
e
c
tor
s
.
I
f
we
manua
ll
y
c
a
lcula
te
us
ing
F
or
mul
a
2,
we
c
a
n
ge
t
the
va
lue
of
C
os
ine
S
im
il
a
r
it
y
is
1
be
c
a
us
e
bo
th
ve
c
tor
s
e
xa
c
tl
y
ha
ve
the
s
a
me
d
ir
e
c
ti
on.
I
n
T
a
ble
2
,
we
ha
ve
two
di
f
f
e
r
e
nt
va
lues
o
f
ve
c
to
r
s
a
nd
we
ge
t
the
C
os
ine
S
im
il
a
r
it
y
s
c
or
e
is
0.
77
74
us
ing
F
or
mul
a
2
.
F
igur
e
11
a
nd
F
igu
r
e
12
a
r
e
the
r
e
s
u
lt
s
that
a
r
e
given
by
ou
r
modul
e
,
C
leve
r
e
e
.
T
hus
,
we
c
a
n
c
onc
lude
that
the
im
pleme
ntation
of
the
C
os
ine
S
i
mi
lar
it
y
method
in
C
leve
r
e
e
is
s
uc
c
e
s
s
.
T
a
ble
1.
F
ir
s
t
s
c
e
na
r
io
to
tes
t
the
im
pleme
ntation
o
f
c
os
ine
s
im
il
a
r
it
y
method
V
a
r
ia
bl
e
C
ondi
ti
on
V
e
c
to
r
di
me
ns
io
n
F
iv
e
V
e
c
to
r
1
[
0,18 0,3
-
0,18 0,49
-
0,18]
V
e
c
to
r
2
[
0,18 0,3
-
0,18 0,49
-
0,18]
E
xpe
c
te
d output
T
he
r
e
s
ul
t
of
t
he
ma
nua
l
c
a
lc
ul
a
ti
on f
or
ve
c
to
r
ma
gni
tu
de
, dot
pr
oduc
t,
a
nd c
os
in
e
s
im
il
a
r
it
y i
s
t
he
s
a
me
a
s
th
e
r
e
s
ul
t
s
how
n i
n F
ig
ur
e
11
.
F
igur
e
11
.
T
he
s
im
il
a
r
it
y
r
e
s
ult
us
ing
c
os
ine
s
im
il
a
r
it
y
method
f
or
the
f
ir
s
t
s
c
e
na
r
io
T
a
ble
2.
S
e
c
ond
s
c
e
na
r
io
to
tes
t
the
im
pleme
ntatio
n
of
c
os
ine
s
im
il
a
r
it
y
method
V
a
r
ia
bl
e
C
ondi
ti
on
V
e
c
to
r
di
me
ns
io
n
T
e
n
V
e
c
to
r
1
[
0,29 0,19
-
0,81 0,59
-
0,44 0,29 0,19
-
0,81 0,59
-
0,44]
V
e
c
to
r
2
[
0,81 0,35
-
0,98 0,18 0,03 0,81 0,35
-
0,98 0,18 0,03]
E
xpe
c
te
d output
T
he
r
e
s
ul
t
of
t
he
ma
nua
l
c
a
lc
ul
a
ti
on f
or
ve
c
to
r
ma
gni
tu
de
, dot
pr
oduc
t,
a
nd c
os
in
e
s
im
il
a
r
it
y i
s
t
he
s
a
me
a
s
th
e
r
e
s
ul
t
s
how
n i
n F
ig
ur
e
1
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
422
-
1
432
1430
F
igur
e
12
.
T
he
s
im
il
a
r
it
y
r
e
s
ult
us
ing
c
os
ine
s
im
il
a
r
it
y
method
f
or
the
s
e
c
ond
s
c
e
na
r
io
3.
2.
E
valu
at
io
n
T
he
e
va
luations
a
r
e
c
a
r
r
ied
out
us
ing
the
T
A
M
e
va
luation
method
a
nd
int
e
r
view
with
thr
e
e
J
a
c
ob
a
dmi
nis
tr
a
tor
s
.
T
he
T
AM
e
va
luation
is
done
us
ing
que
s
ti
onna
ir
e
s
to
mea
s
ur
e
the
P
e
r
c
e
i
ve
d
Us
e
f
ulnes
s
a
nd
P
e
r
c
e
ived
E
a
s
e
of
Us
e
.
F
igur
e
13
pr
e
s
e
nts
t
he
pe
r
c
e
ntage
f
or
e
a
c
h
que
s
ti
on
in
P
e
r
c
e
ived
Us
e
f
ulnes
s
va
r
iable
s
whe
r
e
que
s
ti
on
number
one
(
J
ob
Di
f
f
icult
W
it
hout
)
,
thr
e
e
(
J
ob
P
e
r
f
or
manc
e
)
,
a
nd
f
our
tee
n
(
Us
e
f
ull
f
o
r
S
umm
a
r
y
F
e
a
tur
e
)
ha
ve
the
h
ighes
t
s
c
or
e
.
I
t
is
be
c
a
us
e
C
leve
r
e
e
’
s
f
e
a
tur
e
s
c
a
n
he
lp
a
nd
incr
e
a
s
e
the
pe
r
f
or
manc
e
o
f
the
J
a
c
ob
a
dmi
n
job
,
e
s
pe
c
ially
in
the
Que
s
ti
on
S
umm
a
r
iza
ti
on
f
e
a
tur
e
.
T
he
s
malles
t
s
c
or
e
is
a
c
hieve
d
on
que
s
ti
on
number
twe
lve
a
bout
M
a
ke
s
J
ob
E
a
s
ier
.
I
t
is
due
to
the
f
e
w
us
a
ge
ins
t
r
uc
ti
ons
in
J
a
c
ob
a
dmi
nis
tr
a
tor
pa
ge
s
s
o
s
ometim
e
s
a
dmi
n
f
e
e
ls
c
onf
us
e
d
us
ing
the
a
dmi
n
s
ys
tem.
Ove
r
a
ll
,
the
tot
a
l
pe
r
c
e
ntage
f
or
P
e
r
c
e
ived
Us
e
f
ulnes
s
i
s
79.
17%
(
s
tr
ongly
a
gr
e
e
)
that
thes
e
two
f
e
a
tur
e
s
of
the
C
leve
r
e
e
modul
e
a
r
e
us
e
f
ul.
F
igur
e
14
r
e
pr
e
s
e
nts
the
pe
r
c
e
ntage
f
or
e
a
c
h
qu
e
s
ti
on
in
P
e
r
c
e
ived
E
a
s
e
of
Us
e
va
r
iable
s
whe
r
e
que
s
ti
on
number
s
f
our
tee
n
(
e
a
s
e
to
us
e
f
or
s
umm
a
r
y
f
e
a
tur
e
)
a
nd
f
if
tee
n
(
e
a
s
e
to
us
e
f
or
pa
r
a
phr
a
s
e
f
e
a
tur
e
)
ha
ve
the
highes
t
s
c
or
e
.
I
t
is
be
c
a
us
e
both
f
e
a
t
ur
e
s
make
the
wa
y
to
upda
te
J
a
c
ob's
knowle
dge
e
a
s
ier
.
T
he
s
malles
t
s
c
or
e
is
ob
taine
d
on
que
s
ti
on
n
umber
nine
a
bout
Une
xpe
c
ted
B
e
ha
vior
.
I
t
is
due
to
the
pa
r
a
phr
a
s
ing
r
e
s
ult
s
that
do
not
matc
h
with
t
he
a
c
tual
c
ontext,
thus
making
the
s
ys
tem
unpr
e
dicta
ble.
Othe
r
than
une
xpe
c
ted
be
ha
vior
,
que
s
ti
on
number
s
thr
e
e
(
f
r
us
tr
a
ti
ng)
,
f
our
(
de
pe
nde
nc
e
on
manua
l)
,
a
nd
f
ive
(
menta
l
e
f
f
or
t
)
a
ls
o
ha
ve
the
s
malles
t
s
c
or
e
b
e
c
a
us
e
it
ha
s
a
ne
ga
ti
ve
mea
ning.
P
a
r
a
phr
a
s
ing
r
e
s
ult
s
s
ometim
e
s
give
s
e
mantica
l,
s
yntac
ti
c
a
l,
or
gr
a
m
matica
l
e
r
r
or
s
s
o
a
dmi
ns
mus
t
unde
r
s
tand
the
m
e
a
ning
of
the
s
e
ntenc
e
f
ir
s
t
be
f
or
e
ve
r
i
f
ying
the
r
e
s
ult
s
s
o
it
c
a
us
e
s
the
f
r
us
tr
a
ti
ng
a
nd
menta
l
e
f
f
o
r
t
to
a
dmi
ns
.
De
pe
nde
nc
e
on
M
a
nua
l
a
ls
o
ge
t
s
the
s
malles
t
s
c
or
e
be
c
a
us
e
a
s
a
whole
of
a
s
ys
tem
it
ne
e
ds
a
manua
l
book
s
o
a
dmi
n
can
lea
r
n
how
to
us
e
both
f
e
a
tur
e
s
.
Ove
r
a
ll
,
the
tot
a
l
pe
r
c
e
ntage
f
o
r
P
e
r
c
e
ived
E
a
s
e
of
Us
e
is
72.
57%
(
s
tr
ongly
a
gr
e
e
)
.
B
a
s
e
d
on
the
int
e
r
view
r
e
s
ult
s
c
onduc
ted
to
e
va
lu
a
te
the
C
leve
r
e
e
in
a
qua
li
tative
wa
y,
it
is
known
that
the
pa
r
a
ph
r
a
s
e
ge
ne
r
a
ti
on
gives
good
r
e
s
ult
s
.
How
e
ve
r
,
s
ome
o
f
the
pa
r
a
phr
a
s
e
d
s
e
ntenc
e
s
s
ti
l
l
r
e
quir
e
im
pr
ove
ment.
T
he
r
e
is
a
n
incompr
e
he
ns
ibl
e
s
e
ntenc
e
f
r
om
the
r
e
s
ult
s
of
pa
r
a
phr
a
s
e
be
c
a
us
e
a
wor
d
a
ppe
a
r
s
not
in
the
r
igh
t
plac
e
(
s
yntac
ti
c
a
l
e
r
r
or
)
.
T
hus
,
it
is
im
pa
c
ti
ng
the
c
ontext
a
nd
mea
ning
o
f
the
whole
s
e
ntenc
e
.
T
his
r
e
s
ult
doe
s
not
matc
h
with
the
a
c
tual
c
onte
xt
(
s
e
mantic
e
r
r
o
r
)
.
T
he
c
a
us
e
of
thi
s
pr
oblem
i
s
whe
r
e
s
e
ntenc
e
s
a
nd
wor
ds
a
r
e
not
a
va
il
a
ble
in
the
tr
a
ini
ng
da
tas
e
t.
He
nc
e
,
the
model
ha
s
ne
ve
r
lea
r
ne
d
the
s
e
ntenc
e
s
a
nd
wor
ds
be
f
or
e
.
Anothe
r
r
e
s
ult
of
the
que
s
ti
on
s
umm
a
r
iza
ti
on
f
e
a
tur
e
s
uc
c
e
s
s
f
ull
y
gives
the
de
s
ir
e
d
que
s
ti
on
by
J
a
c
ob’
s
a
dmi
nis
tr
a
tor
s
a
nd
s
hows
the
im
por
tant
que
s
ti
ons
a
s
ke
d
by
us
e
r
s
.
F
igur
e
13
.
T
he
ba
r
c
ha
r
t
f
o
r
the
r
e
s
ult
s
of
pe
r
c
e
ived
us
e
f
ulnes
s
va
r
iable
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
lev
e
r
e
e
:
an
ar
ti
fi
c
ial
ly
int
e
ll
igent
w
e
b
s
e
r
v
ice
for
J
ac
ob
v
oice
c
hatbot
(
Oc
tavany
)
1431
F
igur
e
14
.
T
he
ba
r
c
ha
r
t
f
o
r
the
r
e
s
ult
s
of
pe
r
c
e
ived
e
a
s
e
of
us
e
va
r
iable
s
4.
CONC
L
USI
ON
W
e
pr
opos
e
d
a
nd
de
ve
loped
the
C
lev
e
r
e
e
a
r
ti
f
icia
l
int
e
ll
igenc
e
modul
e
f
or
the
J
a
c
ob
voice
c
ha
tbot
a
ppli
c
a
ti
on.
C
leve
r
e
e
de
li
ve
r
s
two
f
e
a
tur
e
s
:
pa
r
a
p
hr
a
s
e
of
a
ns
we
r
s
a
nd
que
s
ti
ons
s
umm
a
r
iza
ti
on.
T
he
s
e
two
f
e
a
tur
e
s
a
r
e
us
e
f
ul
f
or
upda
ti
ng
J
a
c
ob’
s
knowle
dge
ba
s
e
manua
ll
y
with
the
he
lp
of
a
n
a
dmi
n
is
tr
a
tor
.
T
he
de
ve
lopm
e
nt
of
the
C
leve
r
e
e
modul
e
a
s
a
w
e
b
s
e
r
vice
e
a
s
e
s
the
int
e
gr
a
ti
on
of
the
modul
e
int
o
J
a
c
ob.
T
he
pr
e
-
tr
a
ined
model
of
S
tac
ke
d
R
e
s
idual
L
S
T
M
is
a
ls
o
pr
ove
n
to
be
s
uc
c
e
s
s
f
ul
in
ge
ne
r
a
ti
ng
pa
r
a
phr
a
s
e
of
a
ns
we
r
s
.
I
t
is
a
ls
o
dis
playe
d
in
t
his
s
tudy
that
t
he
model
c
ould
be
us
e
d
to
ge
ne
r
a
te
pa
r
a
phr
a
s
e
ba
s
e
d
on
the
given
tr
a
ini
ng
da
tas
e
t.
T
he
que
s
ti
ons
s
umm
a
r
iza
ti
on
f
e
a
tur
e
powe
r
e
d
by
the
C
os
ine
S
im
il
a
r
it
y
method
with
pr
e
-
tr
a
ined
W
or
d2Ve
c
a
nd
T
e
xtR
a
nk
a
lgor
it
hm
pr
oduc
e
s
s
a
ti
s
f
a
c
tor
y
r
e
s
ult
s
a
s
ve
r
if
ied
by
J
a
c
ob’
s
a
dmi
nis
tr
a
tor
s
.
T
he
T
AM
e
va
luation
method
s
ho
ws
th
a
t
79.
17%
o
f
r
e
s
ponde
nts
s
tr
ongly
a
gr
e
e
tha
t
the
two
f
e
a
tur
e
s
a
r
e
us
e
f
ul
a
nd
72
.
57%
of
r
e
s
ponde
nts
s
tr
ongly
a
gr
e
e
that
the
two
f
e
a
tur
e
s
a
r
e
e
a
s
y
to
us
e
.
RE
F
E
RE
NC
E
S
[1
]
P.
A
.
A
n
g
g
a,
et
al
.
,
"
D
e
s
i
g
n
o
f
c
h
at
b
o
t
w
i
t
h
3
D
av
a
t
ar,
v
o
i
ce
i
n
t
erface,
an
d
faci
a
l
ex
p
res
s
i
o
n
,
"
2
0
1
5
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
S
ci
e
n
ce
i
n
I
n
f
o
r
m
a
t
i
o
n
Tech
n
o
l
o
g
y,
p
p
.
3
2
6
-
3
3
0
,
2
0
1
5
.
[2
]
M.
H
.
Su
,
et
al
.
,
"
A
Ch
at
b
o
t
U
s
i
n
g
L
ST
M
-
b
as
e
d
Mu
l
t
i
-
L
ay
er
E
mb
ed
d
i
n
g
fo
r
E
l
d
erl
y
Care,
"
2
0
1
7
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
O
r
a
n
g
e
Tech
n
o
l
o
g
i
es
(IC
O
T),
p
p
.
7
0
-
74
,
2
0
1
7
.
[3
]
F.
P.
Pu
t
ri
,
H
.
Mei
d
i
a,
an
d
D
.
G
u
n
a
w
an
,
“D
e
s
i
g
n
i
n
g
In
t
e
l
l
i
g
e
n
t
Pers
o
n
a
l
i
ze
d
Ch
at
b
o
t
fo
r
H
o
t
e
l
Serv
i
ce
s
”,
A
CA
I
2
0
1
9
:
P
r
o
cee
d
i
n
g
s
o
f
t
h
e
2
0
1
9
2
nd
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
A
l
g
o
r
i
t
h
m
s
,
Co
m
p
u
t
i
n
g
a
n
d
A
r
t
i
f
i
c
i
a
l
In
t
e
l
l
i
g
e
n
ce
,
p
p
.
4
6
8
-
4
7
2
,
2
0
1
9
.
[4
]
S.
W
i
j
ay
a
an
d
A
.
W
i
ca
k
s
a
n
a,
"
J
aco
b
V
o
i
ce
Ch
at
b
o
t
A
p
p
l
i
cat
i
o
n
U
s
i
n
g
W
i
t
.
ai
fo
r
Pro
v
i
d
i
n
g
In
fo
rma
t
i
o
n
i
n
U
M
N
,
"
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
n
g
i
n
ee
r
i
n
g
a
n
d
A
d
v
a
n
ce
d
Tec
h
n
o
l
o
g
y
,
v
o
l
.
8
,
n
o
.
6
S3
,
p
p
.
1
0
5
-
1
0
9
,
Sep
t
emb
er
2
0
1
9
.
[5
]
R.
Bh
ag
at
a
n
d
E
.
H
o
v
y
,
"
W
h
at
I
s
a
Parap
h
ra
s
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
.
v
o
l
.
3
9
,
n
o
.
3
,
p
p
.
4
6
3
-
4
7
2
,
Sep
t
em
b
er
2
0
1
3
.
[6
]
V
.
A
.
Bh
ag
w
a
t
,
"
D
eep
L
earn
i
n
g
fo
r
Ch
a
t
b
o
t
s
,
"
M.
S.
T
h
es
i
s
,
Sch
.
o
f
Co
m
p
.
Sci
en
ce,
San
J
o
s
e
St
a
t
e
U
n
i
v
.
,
San
J
o
s
e,
CA
,
2
0
1
8
.
[7
]
M.
Su
n
d
ermey
er,
H
.
N
ey
,
an
d
R
Sch
l
ü
t
er,
"
Fro
m
Feed
fo
r
w
ard
t
o
Recu
rre
n
t
L
ST
M
N
eu
ral
N
e
t
w
o
rk
s
fo
r
L
an
g
u
a
g
e
Mo
d
e
l
l
i
n
g
,
"
IE
E
E
/
A
CM
Tr
a
n
s
a
c
t
i
o
n
s
o
n
A
u
d
i
o
,
S
p
eec
h
,
a
n
d
La
n
g
u
a
g
e
P
r
o
ce
s
s
i
n
g
,
v
o
l
.
2
3
,
n
o
.
3
,
p
p
.
5
1
7
-
5
2
9
,
2
0
1
5
.
[8
]
S.
Y
av
u
z,
e
t
al
.
,
"
D
E
E
PC
O
PY
:
G
ro
u
n
d
e
d
Re
s
p
o
n
s
e
G
en
erat
i
o
n
w
i
t
h
H
i
erarch
i
cal
P
o
i
n
t
er
N
et
w
o
r
k
s
,
"
32
nd
Co
n
f
er
e
n
ce
o
n
Ne
u
r
a
l
In
f
o
r
m
a
t
i
o
n
P
r
o
ce
s
s
i
n
g
S
ys
t
em
s
,
M
o
n
t
r
éa
l
,
2
0
1
9
.
[9
]
A
.
Prak
as
h
,
et
a
l
.
,
"
N
eu
ral
Parap
h
ra
s
e
G
en
erat
i
o
n
w
i
t
h
St
ack
e
d
Res
i
d
u
a
l
L
ST
M
N
et
w
o
r
k
s
,
"
Th
e
2
6
th
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
Li
n
g
u
i
s
t
i
cs
,
O
c
t
o
b
er
2
0
1
6
.
[1
0
]
I.
Ferrei
ra,
et
al
.
,
"
Bu
g
Rep
o
rt
Su
mmari
za
t
i
o
n
:
A
n
E
v
al
u
a
t
i
o
n
o
f
Ran
k
i
n
g
T
ec
h
n
i
q
u
es
,
"
2
0
1
6
X
B
r
a
z
i
l
i
a
n
S
ym
p
o
s
i
u
m
o
n
S
o
f
t
wa
r
e
Co
m
p
o
n
en
t
s
,
A
r
ch
i
t
ec
t
u
r
es
a
n
d
R
eu
s
e
,
p
p
.
1
0
1
-
1
1
0
,
2
0
1
6
.
[1
1
]
S.
A.
Bab
ar
a
n
d
P.
D
.
Pat
i
l
,
“
Im
p
ro
v
i
n
g
Perfo
rma
n
c
e
o
f
T
ex
t
Su
mmar
i
zat
i
o
n
,
”
P
r
o
ce
d
i
a
C
o
m
p
u
t
er
S
ci
e
n
ce
,
v
o
l
.
4
6
,
p
p
.
3
5
4
-
3
6
3
,
2
0
1
5
.
[1
2
]
D
.
M.
V
i
c
t
o
r,
e
t
a
l
.
,
"
A
p
p
l
i
cat
i
o
n
o
f
E
x
t
rac
t
i
v
e
T
e
x
t
Su
mmar
i
zat
i
o
n
A
l
g
o
ri
t
h
m
s
t
o
S
p
eech
-
to
-
T
e
x
t
Me
d
i
a,
"
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
H
yb
r
i
d
A
r
t
i
f
i
ci
a
l
In
t
el
l
i
g
en
ce
S
ys
t
e
m
s
,
p
p
.
5
4
0
-
5
5
0
,
2
0
1
9
.
[1
3
]
R.
Mi
h
al
cea
an
d
P.
T
arau
,
"
T
ex
t
Ran
k
:
Bri
n
g
i
n
g
O
r
d
er
i
n
t
o
T
e
x
t
,
"
P
r
o
ceed
i
n
g
s
o
f
t
h
e
2
0
0
4
Co
n
f
er
e
n
ce
o
n
E
m
p
i
r
i
c
a
l
M
e
t
h
o
d
i
n
Na
t
u
r
a
l
La
n
g
u
a
g
e
P
r
o
ces
s
i
n
g
,
p
p
.
4
0
4
-
4
1
1
,
2
0
0
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.