TE
LKOM
NI
KA
Te
le
c
om
munica
tion,
C
omp
u
tin
g,
El
e
ctroni
cs and
Contr
ol
Vo
l.
19
,
No.
3
,
June
2021
,
pp.
911
~
919
IS
S
N: 16
93
-
6930, acc
red
it
ed
First G
ra
de by
Keme
nr
ist
ek
di
kti, D
ec
ree
N
o: 21/E/
KP
T/
2018
DOI: 10.
12
928/
TELK
OMN
I
KA.v1
9i3
.
18877
911
Journ
al h
om
e
page
:
http:
//
jo
ur
nal.
uad.ac
.id
/i
nd
ex.
php/TE
LKOMNIKA
Author i
dentific
ation in
biblio
grap
hic dat
a using d
eep ne
ural
network
s
Fir
da
us
1
, S
iti
Nu
rm
aini
2
, R
eza Firs
an
d
ay
a Ma
li
k
3
,
Ann
isa Darm
awa
hyuni
4
, Muh
am
mad N
aufa
l
Ra
c
hmatull
ah
5
, A
n
dre
Herv
iant Ju
li
ano
6
, T
io A
r
tha
Nugra
ha
7
,
V
arin
do
Ock
ta
Ken
eddi Pu
tra
8
1
, 2, 4
-
8
Inte
ll
i
gent
Sys
te
ms Re
sea
r
c
h
Group,
Univ
er
sita
s Sriwij
aya
,
Pale
mb
ang
,
Indo
nesia
3
Comm
unicati
on
Networks a
nd
I
nforma
t
ion
Secu
rit
y
R
ese
ar
ch
L
a
b,
Univer
si
ta
s Sr
iwij
ay
a
,
Palemb
ang
,
Indon
esia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
15, 2
020
Re
vised
A
ug 23, 2
020
Accepte
d
Aug
31, 202
0
Author
na
me
di
sambi
guation
(AN
D)
is
a
challenging
ta
sk
for
schola
rs
wh
o
mi
ne
b
ibl
iogr
ap
hic
infor
ma
t
ion
for
scie
n
ti
fi
c
knowledge
.
A
construc
t
ive
appr
oac
h
for
r
e
solving
name
a
mbi
guit
y
is
to
use
com
pu
te
r
algorithm
s
to
ide
nti
fy
au
thor
nam
es.
Some
a
l
gorit
hm
-
b
as
ed
d
isam
biguation
m
et
hods
hav
e
bee
n
d
eve
lop
ed
by
com
pu
te
r
a
nd
dat
a
sci
ent
ist
s.
Among
them,
supervise
d
ma
ch
ine
le
arn
in
g
has
be
en
sta
te
d
to
produc
e
decent
to
ver
y
accura
t
e
disam
biguation
result
s.
Th
is
pa
per
pre
sents
a
com
bin
at
ion
o
f
princ
ip
al
com
ponen
t
ana
l
ysis
(PCA
)
as
a
fe
at
ure
r
educ
t
io
n
and
de
ep
n
eur
al
n
et
works
(DN
Ns
),
as
a
su
per
vised
al
gor
ithm
for
c
la
ss
ifying
AN
D
problems.
The
raw
dat
a
is
g
roupe
d
int
o
four
c
la
ss
es,
i.
e
.
,
synony
ms,
homony
ms,
homonym
s
-
synonyms,
and
non
-
homonym
s
-
synonyms
class
ifi
cation.
We
h
av
e
ta
k
en
int
o
ac
coun
t
seve
r
al
hyper
par
a
meter
s
tuni
ng,
such
a
s
le
arn
ing
r
at
e
,
bat
ch
si
ze,
numbe
r
o
f
th
e
neur
on
and
h
id
den
uni
ts,
and
a
nal
yz
ed
the
ir
i
mpa
c
t
on
th
e
ac
cur
ac
y
of
resul
ts.
To
the
best
o
f
our
knowledge,
t
her
e
are
no
pre
v
i
ous
studie
s
with
such
a
sc
hem
e
.
Th
e
pro
posed
DN
Ns
ar
e
validated
wit
h
othe
r
ML
te
chn
ique
s
such
as
Naïve
Bayes
,
ran
dom
for
es
t
(RF),
and
support
vector
ma
ch
ine
(SV
M)
to
produc
e
a
g
ood
class
ifi
er
.
B
y
expl
or
ing
the
result
in
all
dat
a
,
our
proposed
DN
Ns
cl
assifie
r
has
an
ou
tpe
r
forme
d
oth
er
ML
technique
,
with
ac
cur
ac
y
,
pre
ci
sion
,
recal
l
,
and
F1
-
scor
e,
which
is
99.
98
%,
97
.
98%,
97.
86%,
and
99.
99%,
respe
ct
iv
ely.
In
th
e
futur
e,
thi
s
appr
oa
ch
c
a
n
be
e
asily
ext
end
ed
to
any
dat
ase
t
and
any
bibl
iogra
ph
ic rec
ords provi
der
.
Ke
yw
or
d
s
:
Au
t
hor name
di
sambiguat
io
n
Bi
bliog
ra
phic
data
Deep ne
ural
n
e
tworks
Homon
ym
Synon
ym
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Sit
i
Nurmaini
In
te
ll
igent
Syst
ems Resea
rc
h Group
Un
i
ver
sit
as
Sr
i
wija
ya
Pale
mb
a
ng 30
137, I
ndonesi
a
Emai
l:
sit
i_n
ur
maini@
un
s
ri.c
.id
1.
INTROD
U
CTION
Schola
rly
dig
it
al
li
br
aries
pr
ovide
se
rv
ic
es
al
lowing
the
disc
ov
e
r
y
of
mil
li
ons
of
bib
li
ogra
ph
ic
ci
ta
ti
on
record
s,
facil
itati
ng
li
te
ratu
re
researc
h.
T
he
se
re
po
sit
ori
es
co
ntain
a
utho
r
an
d
c
o
-
aut
hors
name,
w
ork
an
d
publica
ti
on
ve
nu
e
,
a
nd
ti
tl
es
of
par
ti
cular
publica
ti
on
s
[
1]
.
A
dig
it
al
li
brar
y
al
s
o
offe
rs
use
fu
l
resea
rc
h
a
nd
dat
a
functi
onal
it
y t
o
hel
p
fun
ding
i
ns
ti
tuti
on
s
gr
a
nt
in
div
i
du
al
s
[
2]
.
Howe
ver,
di
gital
li
br
aries
are
no
t
f
ree
of
error
s
,
su
c
h
as
disp
a
ra
te
cit
at
ion
for
mats, sca
nnin
g, a
nd
data
c
onve
rsion,
am
bigu
ou
s
a
uthor
na
mes, a
nd a
bbre
viati
on
s
of
pu
blica
ti
on
venues
a
nd
ti
tl
es
[
2]
.
Amo
ng
the
er
rors,
the
main
at
te
ntio
n
is
di
rected
to
a
mb
i
guous
a
uthor
names
,
du
e
to
the
dif
ficult
ie
s
in
her
e
nt
i
n
t
he
publica
ti
ons
of
t
he
resea
r
ch
c
ommu
nity
.
It
is
c
h
al
le
nging
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
1693
-
6930
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
,
V
ol.
19
,
No.
3
,
June
20
21
:
91
1
-
919
912
recog
nize
a
pu
blica
ti
on
's
data
owne
d
by
a
n
ind
ivi
dual
,
a
fun
dame
ntal
issue
since
pe
rs
on
al
names
a
r
e
no
t
adequate
ly
d
ist
inct. A
lar
ge n
umber
of
resea
rch
e
rs
a
re c
urr
ently acti
ve
in vari
ou
s
d
isc
i
plines
[3]
.
Au
t
hor
na
me
disam
biguati
on
(
AND)
is
a
c
ru
ci
al
ta
s
k
in
di
gital
li
br
a
ri
es
because
it
c
an
af
fect
th
e
accurac
y
a
nd
qu
al
it
y
of
dig
i
ta
l
li
br
aries
[4]
.
T
ypic
al
ly,
A
ND
issue
s
ma
y
ta
ke
place
in
two
diff
e
re
nt
forms
;
sy
no
nym
a
nd
homon
ym
.
T
he
same
a
utho
r
may
a
ppear
un
der
disti
nct
na
mes
in
the
syn
onym
issue
,
as
they
publish
in
var
i
ou
s
publica
ti
ons
with
va
r
ying
presentat
io
ns
[
5]
.
O
n
t
he
ot
her
hand,
diff
e
ren
t
aut
hors
m
ay
ha
ve
sh
are
d
or
simi
la
r
names
re
fe
rr
e
d
t
o
as
hom
onym
[6]
.
Th
e
syn
onym
a
nd
hom
onym
pro
blems
are
the
major
chall
enges
of
r
ecognizi
ng
the
auth
or
s
hip
of
publica
ti
ons
[
1,
7].
T
hese
may
be
create
d
by
va
rio
us
iss
ues
s
uch
as
misspell
ings,
na
me ch
a
nges
due to
ma
rr
ia
ge, reli
gious
or
ge
nd
e
r
c
onve
rsions, o
r
a
bbre
viati
on
s.
In
rece
nt
yea
r
s,
seve
ral
stu
dies
with
va
riou
s
ap
proac
he
s
ha
ve
bee
n
cond
ucted
t
o
so
lve
A
ND
chall
enges
[6,
8,
9]
.
S
hin
e
t
al.
[
6]
pro
pose
a
c
onve
ntion
al
meth
od
us
in
g
grap
h
framew
ork
f
or
auth
or
disam
biguati
on
, whic
h
res
ol
ved
by grap
h p
ro
ces
ses inclu
di
ng
ver
te
x
(
or
node
)
sp
li
tt
ing
and
mer
ging
ba
sed
on
co
-
a
uthor
sh
i
p.
Yet,
it
is
sti
ll
inade
qu
at
e
i
n
that
minor
c
onditi
ons
s
uch
as
permane
nt
changes
t
o
na
mes
or
aff
il
ia
ti
on
s
suc
h
as
'
a
uthor
prof
il
e
c
ha
ng
e
s'
can
not
be
adequate
ly
ad
dr
ess
ed
.
Li
n
et
al.
[
8]
im
pl
emen
t
hierar
c
hical
a
gglome
rati
ve
cl
us
te
rin
g
f
or
ha
nd
li
ng
the
A
N
D
iss
ue
with
t
wo
at
trib
utes,
i.
e.
,
the
co
-
a
uthors
a
nd
ti
tl
e
a
tt
ribu
te
s.
The
co
-
a
utho
r
'
s
na
me
in
th
e
record
a
re
gro
uped
int
o
cl
us
te
rs
,
an
d
a
con
ce
pt
of
ra
nk
i
ng
confide
nce
to mea
sure t
he
c
onfi
den
ce
of
d
if
fer
e
nt simil
arit
y measu
reme
nt
s is c
reated.
H
us
sai
n a
nd Asg
har
[
9]
us
e a
grap
h
st
r
uctu
ral cl
us
te
ri
ng alg
or
it
hm
di
sassociat
ing a
uthors usi
ng a
gro
up
detect
io
n
al
gorit
hm
a
nd
gr
a
ph
op
e
rati
ons.
Unfortu
natel
y,
it
cannot
detect
hi
gh
ly
a
mb
i
guous
aut
hor
nam
es
in
cases
w
he
re
one
resea
rc
her
ha
s
mu
lt
iple
re
sear
ch
inte
rests.
S
om
e
li
mit
at
ions
that
ca
n
b
e
e
xp
l
or
e
d
i
n
the
fu
t
ur
e
i
nclu
de
sel
f
-
ci
ta
ti
on
s
,
hidden
con
ce
pts
a
nd
e
mail
addresses
of
a
uthor
s.
On
the
oth
e
r
hand,
Ferr
ei
ra
an
d
G
on
çal
ves
[1]
cl
assify
t
he
pu
blica
ti
on
auth
or
s
hip
a
pp
ro
ac
h
int
o
two
typ
es
,
i.e.
aut
hor
gro
up
i
ng
a
nd
aut
hor
assig
nme
nt.
T
he
aut
h
or
gro
up
i
ng
a
ppr
oac
h
cl
us
te
rs
the
a
uthors
base
d
on
t
he
simi
la
r
it
y
of
t
he
publ
ic
at
ion
data
a
tt
ribu
te
[
10,
11]
,
w
hile
the
auth
or
assignme
nt
a
ppr
oac
h
di
rectl
y
assi
gn
s
a
pu
blica
ti
on
to
th
e
auth
or
by
buil
ding
a
mod
el
that
represe
nts
the
auth
or
[
12, 13]
.
This
pap
e
r
highli
gh
ts
the
a
ut
hor's
as
sig
nm
e
nt
ty
pe
to
rec
ognize
publica
ti
on
aut
hors
hip
.
In
t
he
ty
pe
,
there ar
e t
wo
a
ppr
oach
es t
o
le
arn
i
ng
; cl
assifi
cat
ion
and clus
te
ring. Th
e a
dv
antage
of
the c
la
ssific
at
ion
m
et
hod
is
it
s
ef
ficacy
wh
e
n
face
d
with
man
y
ci
ta
ti
on
ex
a
mp
le
s
f
or
e
ach
a
utho
r.
I
n
c
on
t
rast,
t
he
cl
ust
ering
met
hod
needs
pr
i
vileged
i
nfo
rmati
on
ab
out
the
ap
pro
pr
ia
te
numb
e
r
of
a
uthors
or
t
he
num
ber
of
a
uthor
c
la
sses
and
may
ta
ke
so
me
ti
me
t
o
determi
ne
t
heir
par
a
mete
rs
[
1]
.
Some
rese
arch
e
rs
us
e
d
t
he
a
utho
r
a
ssignment
a
ppro
a
ch
wit
h
cl
assifi
cat
ion
[
12, 1
4,
15]
.
Sti
ll
, th
e
resu
lt
s
a
re
no
t
sati
sfy
i
ng i
n
F
1
-
sco
re a
nd acc
ur
ac
y
[
12,
14]
. T
he
u
se
of t
he
arti
fici
al
ne
ur
al
net
wor
ks
ap
proach
is
al
rea
dy
exp
l
or
e
d
to
rec
ognize
the
aut
hors
hip
of
the
publica
ti
on.
H
oweve
r,
the
perf
orman
ce
gets
poor
r
ecal
l
with
a
go
od
re
su
lt
on
accurac
y
[15]
.
T
o
en
hance
the
performa
nce
of
conve
ntion
al
ne
ur
al
netw
ork
al
gorithms,
a
DNNs
with
m
ulti
ple
la
yer
s
i
s
pro
posed
i
n
this
resea
rch.
DNNs
hav
e
a
stron
g
abili
ty
to
feat
ur
e
le
ar
ning
i
n
ma
ny
ta
sk
s
a
nd
s
olv
e
the
publica
ti
on
a
uthorship
prob
le
m
[4]
.
DNNs
ca
n
buil
d
a
gen
e
ral
m
od
el
t
hat
co
ul
d
disam
biguate
auth
or
na
me
on
a
ste
p
-
by
-
ste
p
ba
sis
w
he
n
new
publica
ti
on
rec
ords
a
re
inte
grat
ed
into
the
da
ta
set
.
This
pa
per
al
s
o
e
xplo
res
f
our
co
mb
i
nations
of
t
ypes
of
publica
ti
on
s
da
ta
(m
ulti
cl
ass
cl
assifi
cat
ion)
p
roblems
in
the
cl
assifi
cat
io
n
ta
s
k,
i.e.
,
s
ynon
ym
s
,
homonyms,
homon
ym
s
-
s
ynon
ym
s
,
an
d
non
-
homon
ym
s
-
sy
no
nyms
cl
as
sific
at
ion
,
w
hich
ar
e
the
mai
n
prob
le
ms
of
auth
or
identific
at
ion
[
16]
.
F
or
co
mpa
rison
s
,
Naïve
Ba
yes,
r
an
dom
forest
,
S
V
M
a
re
us
e
d
f
or
be
nc
hm
a
rk
i
ng
t
he
resu
lt
of cla
ssifie
r pe
rformance
.
2.
RESEA
R
CH MET
HO
D
In
t
his p
ape
r,
t
he
met
hod
f
or
au
th
or
assig
nme
nt
meth
od
is
t
hro
ugh
assi
gn
i
ng
a
re
fere
nce
to
a
s
pecific
auth
or
by
buil
ding
a
m
od
el
t
hat
represe
nts
the
auth
or
us
i
ng
t
he
cl
assifi
cat
ion
te
ch
nique
[
1]
.
A
publ
ic
at
ion
dataset
will
ha
ve
f
our
dif
fere
nt
case
s;
homonym,
s
ynonym
,
s
ynon
ym
-
ho
monym
,
a
nd
non
-
s
ynonym
-
hom
onym.
A
hom
onym
is
the
cases
wh
e
n
diff
e
re
nt
per
s
ons
sh
a
re
t
he
sa
me
name,
a
nd
s
ynonym
is
the
cases
wh
e
n
the
name
of
a
pa
rtic
ular
auth
or
is
give
n
in
se
ver
al
dif
f
eren
t
wa
ys
[
1]
.
T
he
s
ynon
ym
-
hom
onym
cas
e
mea
ns
the
sa
mp
le
data
ha
s
both
a
synon
ym
a
nd
homon
ym
ca
se.
Ot
herwi
se,
the
non
-
s
ynon
ym
-
hom
onym
case
m
us
t
not
hav
e
a
sy
no
nym
or ho
monym
case
.
This
pa
pe
r
pr
opos
e
d
the
a
ut
hor
identific
a
ti
on
pro
cessi
ng
that
co
ns
ist
s
of
four
sta
ge
s;
(i)
data
pr
e
par
at
io
n,
(ii)
featu
re
ext
rac
ti
on
,
(iii
)
cl
assi
ficat
ion
,
a
nd
(i
v)
performa
nc
e
evalu
at
io
n
(s
ee
in
Fig
ur
e
1).
The
dig
it
al
bib
li
ogr
aph
ic
&
li
brar
y
project
(
DB
LP)
la
bele
d
da
ta
set
is
implemente
d
i
n
this
study.
I
n
the
featur
e
extracti
on,
the
new
feat
ur
es
a
r
e
extra
ct
ed
fro
m
eac
h
at
trib
ut
e
in
a
dataset
.
Wh
il
e
i
n
t
he
cl
assifi
cat
ion
,
th
e
proce
ss
a
nd
le
a
rn
th
os
e
featur
es
t
o
re
present
the
s
pec
ific
auth
or
s
.
T
he
co
mp
a
rison
of
tw
o
cl
assifi
ers,
S
V
M
an
d
DNN,
has
been
ex
plored
i
n
this
st
udy
.
In
t
he
la
st,
th
e
cl
assifi
ers
will
be
evaluate
d
with
fi
ve
performa
nce
metri
cs
(i.e.
,
accurac
y, sen
sit
ivit
y,
s
pecific
it
y,
pr
eci
sio
n, a
nd F
1
-
sc
ore) t
o vali
date t
he pr
opos
e
d mo
del.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
Author i
den
ti
fi
cation i
n bi
blio
gr
ap
hic
da
t
a
usi
ng
dee
p ne
ural netw
or
ks
(
Fi
rd
aus
)
913
F
e
a
t
u
re
E
x
t
ra
c
t
i
on
C
l
a
s
s
i
f
i
c
a
t
i
on
P
e
rf
or
m
a
n
c
e
E
v
a
l
ua
t
i
on
D
a
t
a
P
re
pa
ra
t
i
on
Figure
1. A
uthor
ide
ntific
at
ion
processi
ng stage
2
.
1
.
D
ata
pre
p
ar
at
i
on
In
t
his
pap
e
r,
we
im
plement
the
a
uthor
na
me
disa
mb
i
guat
ion
la
bele
d
da
ta
gen
e
rated
by
D
r.
Gile
s
researc
h
la
b
at
the
Pe
nnsy
l
va
nia
Stat
e
U
nive
rsity
[
17,
18]
,
an
d
cl
eane
d
by
Kim
[
19]
.
T
he
cl
ea
ning
pr
ocess
resu
lt
ed
in
5018
name
insta
nc
es
with
480
di
sti
nct
auth
o
r
s
(au
t
hor
la
bels)
an
d
456
disti
nct
pr
ese
nted
names
wh
e
re
eac
h
a
ut
hor
has
1
to
48
0
ref
e
ren
ce
s.
T
he
dataset
co
m
pr
ise
s
the
a
uthor
’s
pr
e
sente
d
name,
a
uthor
l
abel,
auth
or
s
name
,
venue,
a
nd
ti
tle.
The
la
bel
of
four
A
ND
pro
blems
are
unav
ai
la
ble
in
the
da
ta
set
.
T
her
e
fore,
w
e
need
cat
e
goriz
ing
these
four
cases
with
(
1
-
4),
f
or
hom
onym,
syn
onym,
hom
onym
-
sy
no
nym,
a
nd
no
n
-
homon
ym
-
syn
onym,
resp
ect
i
vely;
ℎ
=
→
1
↦
,
≥
2
(1)
=
→
1
↦
,
≥
2
(2)
W
he
re
is
pres
ented
na
me,
is
aut
hor,
is
the
num
ber
of
,
a
nd
is
the
num
ber
of
.
For
t
he
hom
onym
case,
one
has
the
num
be
r
of
more
tha
n
or
e
qu
al
to
2,
w
herea
s,
f
or
the
syn
onym
ca
se,
on
e
has
the
num
ber
of
mo
re
tha
n o
r
e
qu
al
t
o 2.
ℎ
=
ℎ
∩
(3)
ℎ
=
(
ℎ
∪
)
(4)
2
.
2
.
Fe
at
ure
e
xtracti
on
The
featu
re
e
xt
racti
on
for
a
ut
hor
ide
ntific
at
ion
ca
n
be
pre
sented
in
Fi
gu
re
2.
Fi
gure
2
show
s
the
pr
e
processi
ng
ph
a
se
f
or
aut
hor
i
den
ti
ficat
io
n
,
data
normal
iz
at
ion
,
featur
e
extracti
on,
fea
tures
c
oncat
en
at
ion
,
and
featu
res
r
e
du
ct
io
n.
The
fe
at
ur
es
that
bec
om
e
the
cl
as
sif
ie
r
input
are
e
xt
racted
f
rom
da
ta
set
at
tribu
te
s.
The
dataset
at
trib
utes
c
onsist
of
t
w
o
ty
pes
of
at
tri
bu
te
s;
cat
e
gori
cal
(prese
nted
name,
aut
hor
na
me,
venue,
an
d
ti
tl
e)
and
nume
rical
(year
).
Ca
te
gorical
at
trib
ute
s
are
proces
se
d
int
o
a
one
-
hot
nume
ric
ar
r
ay,
w
hile
num
erical
at
tribu
te
s
are
l
eft
as
is.
The
f
irst
featur
e
gro
up
is
e
x
tract
ed
from
the
pr
es
ented
name
at
trib
ute.
Thes
e
f
eat
ur
e
s
are
e
xtracted
by
c
reati
ng
a
on
e
-
hot
numeric
arr
a
y
of
la
bels
encodin
g
disti
nct
prese
nted
na
mes.
T
he
enc
od
i
ng
la
bel
is
the
c
onver
si
on
of
cat
egorical
da
ta
int
o
numerical
.
T
his
process
pr
oduce
s
a
feat
ur
e
in
the
f
orm
of
a
de
ns
e
bin
a
ry
a
rr
a
y
with
an
a
rr
a
y
le
ngth
e
qu
al
t
o
th
e
numb
e
r
of
di
sti
nct
pr
ese
nte
d
na
mes.
The
s
econd
featu
re
gro
up
is
extracte
d
f
rom
the
aut
hors
na
me
at
trib
ute.
F
or
t
he a
uthors
name a
tt
ri
bu
te
,
a
la
bel
is
e
ncoded
f
or
al
l
the
disti
nct
auth
o
rs
names
create
d
i
n
t
he
same
w
ay
as
the
pr
ese
nted
name.
T
he
n,
only
yea
r
at
trib
utes
a
re
sp
eci
f
ic
al
ly
normali
zed
w
it
h
a
min
-
ma
x
sc
al
ing
in
(
5).
=
−
−
(5)
wh
e
re
is t
he
i
nput,
is t
he
scal
ed
in
put,
is t
he mi
nimum
valu
e, and
is t
he
m
aximum
value
.
The
t
hir
d
group
is
t
he
ve
nu
e
a
tt
ribu
te
,
w
hich
has
t
he
sa
me
process
us
e
d
on
t
he
pr
ese
nte
d
na
me.
Un
li
ke
in
oth
e
r
gro
ups,
th
ere
are
tw
o
main
pre
pro
cessi
ng
sta
ges
for
ti
tl
e
at
trib
ut
e
in
t
he
te
xt
a
tt
ribu
te
s
su
c
h
as
te
xt
normali
zat
ion
and
feat
ur
e
e
xt
racti
on. I
n
a
ddi
ti
on
,
f
or
te
xt
nor
mali
zat
ion
,
l
emmat
iz
at
ion
a
nd
la
ncaster
st
emmer
are
use
d,
wh
il
e
for
featu
re
e
xtracti
on,
te
r
m
fr
e
quenc
y
-
in
ve
rse
do
c
um
e
nt
fr
e
quenc
y
(TF
-
I
DF
)
,
,
is
use
d
in (
6).
(
,
,
)
=
(
,
)
∙
(
,
)
(6)
wh
e
re
(
,
)
is
the
te
r
m
f
reque
ncy,
t
he
numb
e
r
of
t
ime
that
te
rm
oc
cur
s
in docum
ent
,
wh
ic
h
do
c
um
e
nt
in
corp
us
,
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
1693
-
6930
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
,
V
ol.
19
,
No.
3
,
June
20
21
:
91
1
-
919
914
P
re
s
e
n
t
e
d
n
a
m
e
A
u
t
h
o
r
s
n
a
m
e
Y
e
a
r
V
e
n
u
e
T
i
t
l
e
N
o
r
m
a
l
i
z
e
d
T
e
x
t
D
u
m
m
y
v
a
r
i
a
b
l
e
D
u
m
m
y
v
a
r
i
a
b
l
e
D
u
m
m
y
v
a
ri
a
b
l
e
TF
-
I
D
F
T
e
x
t
N
o
rm
a
l
i
z
a
t
i
o
n
M
i
n
-
m
a
x
s
c
a
l
l
e
r
n
o
r
m
a
l
i
z
a
t
i
o
n
P
C
A
R
e
d
u
c
e
d
f
e
a
t
u
r
e
s
v
e
c
t
o
r
C
l
a
s
s
i
f
i
e
r
0
t
o
1
s
c
a
l
e
f
e
a
t
u
r
e
D
a
t
a
N
o
r
m
a
l
i
z
a
t
i
o
n
F
e
a
t
u
r
e
s
Ex
t
r
a
c
t
i
o
n
F
e
a
t
u
r
e
s
R
e
d
u
c
t
i
o
n
o
n
e
-
h
o
t
n
u
m
e
r
i
c
a
rr
a
y
o
n
e
-
h
o
t
n
u
m
e
r
i
c
a
r
r
a
y
o
n
e
-
h
o
t
n
u
m
e
ri
c
a
r
r
a
y
F
e
a
t
u
r
e
s
C
o
n
c
a
t
i
n
a
t
i
o
n
F
e
a
t
u
re
s
c
o
n
c
a
t
i
n
a
t
i
o
n
P
r
e
p
r
o
c
e
s
s
i
n
g
Figure
2. A
uthor
Identific
at
io
n pr
e
processi
ng
ph
a
se
In
t
he
featu
re
con
cat
e
natio
n,
the
extracte
d
featur
e
s
are
c
ombine
d
from
t
he
prese
nted
na
me,
aut
hor
la
bel,
aut
hors
na
me,
year
,
ve
nue,
a
nd
ti
tl
e
resu
lt
s
in
ma
ny
fe
at
ur
es.
S
om
e
a
uthor a
ssig
nm
e
nt
stu
dies
ha
ve
us
e
d
the
blo
c
king
method
befor
e
c
onduct
in
g
al
gorith
m
ic
disam
biguati
on
ta
s
ks
f
or
reducin
g
co
mputi
ng
ti
me
[
20,
21]
.
Dissimi
la
r
t
o
them,
t
his
pa
pe
r
reduces
the
di
mensional
it
y
of
feat
ur
e
s
via
pri
nci
pal
c
ompone
nt
analysis
(
PCA
)
.
PC
A
us
e
s
a
n
or
t
hogonal
tra
nsfo
rmati
on
to
c
hange
a
set
of
ob
s
er
vations
of
va
riables
c
orre
la
te
d
to
the
value
of
var
ia
bles
that
a
re
no
t
li
near
l
y
correla
te
d,
cal
le
d
t
he
pri
ncipa
l
com
pone
nt
[
22]
.
Han
et
al.
pro
pose
PCA
f
or
A
N
D
pro
blems
beca
us
e
of
it
s
a
bili
ty
t
o
rem
ov
e
th
e
li
near
co
rr
el
a
ti
on
s
a
nd
im
p
r
ov
e
the
gen
e
ral
iz
at
ion
performa
nce
[
23]
.
I
n
this
resea
rch,
we
fine
-
t
une
the
numb
e
r
of
featu
res
f
rom
2
to
2500
fe
at
ur
es
to
fin
d
the
bes
t
cl
assifi
er p
e
rfo
rma
nce.
2
.
3
.
Cl
as
sific
ati
on
The
cl
assifi
e
r
gets
in
pu
t
fro
m
ext
racted
fe
at
ur
es
i
n
the
previ
ou
s
proc
es
s.
T
he
cl
assifi
er
le
ar
ns
the
featur
e
s
of
the
trai
ning
dataset
to
de
te
rmine
a
r
efere
nce
t
o
a
spe
ci
fic
aut
hor.
I
n
this
pa
pe
r,
t
he
pro
pose
d
cl
as
sifie
r
was
c
onduct
ed
us
in
g
D
NNs
c
la
ssifie
rs
a
nd
Naïve
Ba
yes,
r
andom
f
or
e
st
,
SVM
as
co
mpa
risons.
The
pr
opos
e
d
cl
assifi
er
us
es
DNNs
-
base
d
c
la
ssifie
r.
DNN
s
re
fer
to
ne
ural
netw
orks
wi
th
a
la
r
ge
num
ber
of
hidde
n
l
ayer
s
.
With
dee
p
a
rc
hitec
ture
in
Ne
ur
al
Net
works,
D
N
Ns
ca
n
re
present
hi
gh
e
r
c
omplexit
y
f
un
c
ti
on
s.
T
his
a
bili
ty
is
po
s
sible
by inc
reasin
g
the
nu
mb
e
r of
la
yer
s
and n
e
uro
ns
i
n t
he
la
ye
r
[
24]
(
Figure
3).
Figure
3. Pro
pose
d
D
N
Ns
a
rc
hitec
ture
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
Author i
den
ti
fi
cation i
n bi
blio
gr
ap
hic
da
t
a
usi
ng
dee
p ne
ural netw
or
ks
(
Fi
rd
aus
)
915
In
t
his r
esea
rc
h, D
NNs train
w
it
h
cat
egorical
cro
ss
-
e
ntr
opy
l
os
s
functi
on
as
in
(
7)
a
nd
rect
ifie
d
li
near
un
it
(ReL
U)
as
in
(
8)
as
act
iva
ti
on
functi
ons.
Thr
ee
hidden
l
ayer
s
with
50,
100,
150,
a
nd
200
ne
uro
ns
,
w
it
ho
ut
dro
pout,
an
d
0.1
to
0.5
dr
opout
values
a
rc
hitec
ture
are
us
e
d
in
this
e
xp
e
riment
to
get
opti
mu
m
cl
assifi
cat
ion
performa
nce
(
Table
1
).
T
he
par
a
mete
rs
tha
t
pr
od
uce
the
best
cl
assifi
cat
ion
performa
nc
e
are
sel
ect
ed
as
the
neural
netw
ork
buil
der
pa
ram
et
erser
,
w
hich
is
the
in
put
vec
tor
,
is
the
de
sir
ed
ou
t
pu
t,
an
d
ẏ
is
the
pre
dicte
d
ou
t
pu
t.
W
e
tr
y 100 e
poch
s, 0.
01 lear
ning r
at
e an
d 64 batc
h si
ze for all
sce
nar
i
os
.
(
y
,
y
̂
)
(
y
,
y
̂
)
=−
∑
∑
(
y
ij
∙
log
(
y
̂
ij
)
)
N
i
=
0
M
j
=
0
∑
∑
(
y
ij
∙
log
(
y
̂
ij
)
)
N
i
=
0
M
j
=
0
(7)
(
x
)
(
x
)
=
{
0
,
x
≤
0
x
,
x
>
0
{
0
,
x
≤
0
x
,
x
>
0
(8)
Table
1.
D
NN
s
arc
hitec
ture
a
nd tu
ning
par
a
mete
rs
Layer
Nu
m
b
er
o
f
Neuron
s
Activ
atio
n
Fun
ctio
n
Inp
u
t
PCA g
en
erate
d
-
Hid
d
en
layer 1
5
0
,
1
0
0
,
1
5
0
,
2
0
0
ReLU
Hid
d
en
layer 2
5
0
,
1
0
0
,
1
5
0
,
2
0
0
ReLU
..
…
…
Hid
d
en
layer 8
5
0
,
1
0
0
,
1
5
0
,
2
0
0
ReLU
Ou
tp
u
t layer
266
So
ftm
ax
2
.
4
.
Ev
alu
at
io
n
The
dataset
is
div
ide
d
by
80
%
of
t
rainin
g
da
ta
and
remain
ing
for t
he t
est
ing
data.
Be
f
ore
sp
li
tt
ing
,
by
consi
der
i
ng
t
he
numb
e
r
of
a
uthor
ref
e
ren
c
es
and
t
he
dist
rib
ution
of
trai
ning
an
d
te
sti
ng
data,
we
re
moved
auth
or
s
w
ho
ha
ve
le
ss
tha
n
fiv
e
ref
e
ren
ce
s
f
r
o
m
th
e
dataset
,
so
the
num
ber
of
data
dec
reas
es
from
5018
t
o
44
19
name
i
ns
ta
nce
s
. T
he
c
omparis
on of t
he
ra
w d
at
aset
an
d t
he pre
par
e
d datas
et
is p
re
sente
d
in Ta
ble 2.
The
numb
e
r
of
rest
aut
hor
det
ermines
the nu
mb
e
r
of
cl
asse
s
us
e
d
in
the
cl
assifi
cat
ion
.
T
her
e
fore,
t
he
numb
e
r
of
cl
as
ses
set
as
266
cl
asses
of
a
uthors.
Table
3
pr
esents
the
rec
ord
numb
e
r
a
nd
portion
of
eac
h
f
our
AND
pr
ob
le
m
aff
ect
ed
by
dat
a
cl
eani
ng
a
nd
sp
li
tt
ing
.
Data
cl
eaning
giv
e
s
a
fai
r
e
ff
ect
on
the
portio
n
of
fou
r
AND pr
oble
ms
, while
d
at
a
s
pl
it
ti
ng
does
not
aff
ect
.
Table
2.
C
omp
os
it
ion o
f raw
dataset
compa
r
ed
to
the
prepa
red dataset
Raw d
ataset
Prepared d
ataset
Nam
e
ins
tan
ces
5018
4419
Distin
ct auth
o
rs
480
266
Distin
ct presen
ted
n
am
es
456
303
Distin
ct ven
u
es
1004
923
Distin
ct
co
-
au
th
o
r
n
am
es
4653
3733
Year
rang
e
1959
-
2
0
1
0
1959
-
2
0
1
0
Sy
n
o
n
y
m
auth
o
rs/row af
fe
cted
4
6
/1
0
6
9
4
6
/1
1
2
0
Ho
m
o
n
y
m
presen
t
ed
names/row
af
f
e
cted
6
2
/7
8
7
1
5
/3
2
8
No
n
-
sy
n
o
n
y
m
-
h
o
m
o
n
y
m
r
o
w
affect
ed
2988
2861
Sy
n
o
n
y
m
-
h
o
m
o
n
y
m
r
o
w
af
fecte
d
174
110
Table
3.
C
omp
os
it
ion o
f
trai
ni
ng
a
nd test
in
g datase
t
AND P
rob
lem
Raw d
ataset
Prepared d
ataset
Tr
ain
in
g
datas
et
Testin
g
datas
et
Reco
rd
n
u
m
b
er
(%)
Reco
rd
n
u
m
b
er
(%)
Reco
rd
n
u
m
b
er
(%)
Reco
rd
n
u
m
b
er
(%)
Sy
n
o
n
y
m
1069
2
1
.30
%
1120
2
5
.34
%
900
2
5
.46
%
221
2
5
.00
%
Ho
m
o
n
y
m
787
1
5
.70
%
328
7
.42
%
262
7
.41
%
66
7
.46
%
Sy
n
o
n
y
m
-
Ho
m
o
n
y
m
174
3
.46
%
110
2
.50
%
85
2
.40
%
25
2
.83
%
No
n
-
Sy
n
o
n
y
m
-
Ho
m
o
n
y
m
2988
5
9
.54
%
2861
6
4
.74
%
2288
6
4
.72
%
572
6
4
.70
%
Total
5018
1
0
0
.00
%
4419
1
0
0
.00
%
3535
1
0
0
.00
%
884
1
0
0
.00
%
We
a
pp
li
ed
fe
w
sta
ti
sti
c
methods
t
o
eval
uat
e
the
pe
rfo
rma
nce
of
pro
po
se
d
met
hods
,
su
c
h
as
a
ver
a
ge
accurac
y
as
in
(
9)
,
preci
sio
n
as
in
(
10),
re
cal
l
as
in
(
11)
,
an
d
F
1
-
sco
re
as
in
(12
)
to
e
valuate
our
m
et
hod
performa
nce
[25]
.
=
∑
+
+
+
+
+
=
1
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
1693
-
6930
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
,
V
ol.
19
,
No.
3
,
June
20
21
:
91
1
-
919
916
=
∑
+
=
1
(10)
=
∑
+
=
1
(11)
1
−
=
2
∙
∙
+
(12)
W
he
re
is
tr
ue
po
sit
ive
,
is
tr
ue
ne
gative
,
is
false
posit
ive,
is
false
ne
gativ
e,
a
nd
is
the
num
ber
of
cl
asses.
T
he
pe
rformance
e
va
luati
on
meth
od
is
ca
rr
ie
d
ou
t
on
f
our
a
uthor
na
me
di
sambi
gu
at
io
n
issues;
homon
ym
,
s
ynonym,
non
-
syn
onym
-
homon
ym, a
nd s
ynonym
-
homon
ym
.
3.
RESU
LT
S
A
ND
DI
SCUS
S
ION
Th
e
feat
ur
e
e
xtracti
on
process
ge
ne
rates
a
numb
e
r
of
featu
r
es
f
or
eac
h
at
tr
ibu
te
.
7793
fea
tures
we
re
gen
e
rated
w
it
h detai
ls
pr
ese
nt
ed
i
n
Ta
ble
4.
The n
umber
ca
n
i
ncr
ease
the
com
pu
ta
ti
onal
cost.
Th
us
,
it
re
du
ce
d
by
ap
ply
i
ng
P
CA
become
1382
featu
res.
To
obta
in
a
n
op
ti
m
um
cl
ass
ific
at
ion
perfor
mance,
var
io
us
D
NNs
structu
res
in
t
he
le
arn
in
g
pr
oc
ess
wer
e
e
xami
ned.
The
224
st
ru
ct
ur
es
wer
e
de
te
rmin
e
d
an
d
validat
ed
before
the
sel
ect
ion
of
th
e
best
model.
All
cl
assifi
ers
wer
e
ar
ra
ng
e
d
in
tw
o
pr
oc
esses
i.e.,
trai
ning
an
d
te
sti
ng.
All
processi
ng ti
me r
es
ult o
f 224
DNNs st
ru
ct
ures is p
rese
nted i
n
Ta
ble
5
.
Table
4.
N
umb
er
of
featu
res.
Attribu
te
Featu
re
Extractio
n
Techn
iq
u
e
Nu
m
b
er
o
f
f
eatu
r
e
s
Presen
ted
name
Du
m
m
y
va
riable
303
List of auth
o
r's na
m
es
Du
m
m
y
v
a
riable
3733
Year
No
n
e
1
Ven
u
e
Du
m
m
y
va
riable
923
Title
TF
-
I
DF
2833
Table
5
. M
od
el
accu
racy (%
) f
or 24
4
D
N
Ns
s
tructu
res
Neu
ron
Lear
n
in
g
Rate
Hid
d
en
L
ay
er
1
2
3
4
5
6
7
8
50
1E
-
01
9
9
,55
3
5
9
9
,30
6
8
9
9
,28
5
5
9
9
,26
6
0
9
9
,26
9
4
9
9
,26
6
0
9
9
,26
9
4
9
9
,26
5
1
50
1E
-
02
9
9
,98
0
4
9
9
,97
7
0
9
9
,94
8
1
9
9
,91
0
7
9
9
,78
9
1
9
9
,72
2
7
9
9
,65
3
8
9
9
,51
0
9
50
1E
-
03
9
9
,98
1
3
9
9
,97
9
6
9
9
,97
6
2
9
9
,97
0
2
9
9
,95
9
2
9
9
,92
9
4
9
9
,90
9
8
9
9
,86
2
2
50
1E
-
04
9
9
,79
6
7
9
9
,85
3
7
9
9
,91
9
2
9
9
,93
3
7
9
9
,88
5
2
9
9
,85
2
9
9
9
,83
1
6
9
9
,72
7
0
50
1E
-
05
9
9
,39
2
7
9
9
,31
0
2
9
9
,26
9
4
9
9
,28
3
8
9
9
,26
2
6
9
9
,28
4
7
9
9
,26
4
3
9
9
,27
2
8
50
1E
-
06
9
9
,25
2
4
9
9
,26
0
9
9
9
,25
2
4
9
9
,26
0
0
9
9
,25
8
3
9
9
,26
8
5
9
9
,26
6
8
9
9
,26
7
7
50
1E
-
07
9
9
,25
4
1
9
9
,24
9
8
9
9
,25
3
2
9
9
,25
0
7
9
9
,24
9
8
9
9
,25
1
5
9
9
,24
8
1
9
9
,24
9
8
100
1E
-
01
9
9
,73
5
5
9
9
,30
4
3
9
9
,26
9
4
9
9
,26
2
6
9
9
,25
4
1
9
9
,26
6
0
9
9
,26
9
4
9
9
,26
9
4
100
1E
-
02
9
9
,98
1
3
9
9
,97
7
9
9
9
,95
4
1
9
9
,84
6
1
9
9
,71
1
7
9
9
,59
2
6
9
9
,41
9
1
9
9
,38
0
8
100
1E
-
03
9
9
,98
3
0
9
9
,98
2
1
9
9
,97
7
9
9
9
,97
7
9
9
9
,96
6
0
9
9
,96
3
4
9
9
,95
4
1
9
9
,95
6
6
100
1E
-
04
9
9
,95
1
5
9
9
,97
6
2
9
9
,97
5
3
9
9
,97
3
6
9
9
,95
7
5
9
9
,95
2
4
9
9
,92
2
6
9
9
,87
0
7
100
1E
-
05
9
9
,47
1
8
9
9
,28
7
2
9
9
,28
8
9
9
9
,28
1
3
9
9
,32
3
8
9
9
,28
0
4
9
9
,34
8
5
9
9
,29
0
6
100
1E
-
06
9
9
,25
4
1
9
9
,24
9
8
9
9
,25
3
2
9
9
,25
6
6
9
9
,25
5
8
9
9
,26
3
4
9
9
,26
2
6
9
9
,26
3
4
100
1E
-
07
9
9
,24
9
8
9
9
,25
1
5
9
9
,24
9
8
9
9
,24
9
0
9
9
,24
9
8
9
9
,25
0
7
9
9
,25
1
5
9
9
,25
2
4
150
1E
-
01
9
9
,78
9
1
9
9
,29
5
8
9
9
,27
2
8
9
9
,26
9
4
9
9
,26
7
7
9
9
,26
9
4
9
9
,25
6
6
9
9
,25
5
8
150
1E
-
02
9
9
,97
7
0
9
9
,97
7
9
9
9
,94
3
9
9
9
,74
5
7
9
9
,60
7
9
9
9
,47
1
8
9
9
,43
3
5
9
9
,33
4
9
150
1E
-
03
9
9
,98
3
8
9
9
,98
2
1
9
9
,98
0
4
9
9
,97
6
2
9
9
,97
6
2
9
9
,96
6
0
9
9
,96
0
0
9
9
,93
9
6
150
1E
-
04
9
9
,97
5
3
9
9
,98
3
0
9
9
,98
0
4
9
9
,97
8
7
9
9
,97
7
0
9
9
,95
8
3
9
9
,94
0
5
9
9
,91
4
9
150
1E
-
05
9
9
,50
4
1
9
9
,32
3
0
9
9
,34
0
8
9
9
,36
4
6
9
9
,37
4
9
9
9
,36
8
9
9
9
,39
1
0
9
9
,35
7
0
150
1E
-
06
9
9
,26
0
0
9
9
,26
1
7
9
9
,25
6
6
9
9
,25
7
5
9
9
,25
7
5
9
9
,26
2
6
9
9
,26
3
4
9
9
,27
9
6
150
1E
-
07
9
9
,25
1
5
9
9
,25
0
7
9
9
,25
4
1
9
9
,25
6
6
9
9
,24
9
0
9
9
,25
0
7
9
9
,25
2
4
9
9
,24
9
8
200
1E
-
01
9
9
,81
8
0
9
9
,27
2
8
9
9
,26
9
4
9
9
,26
6
0
9
9
,26
6
0
9
9
,26
9
4
9
9
,26
9
4
9
9
,26
0
9
200
1E
-
02
9
9
,97
5
3
9
9
,97
7
9
9
9
,92
0
9
9
9
,60
3
6
9
9
,58
6
6
9
9
,37
0
6
9
9
,34
0
0
9
9
,27
6
2
200
1E
-
03
9
9
,98
3
8
9
9
,98
3
0
9
9
,98
0
4
9
9
,97
5
3
9
9
,97
4
5
9
9
,97
1
1
9
9
,96
4
3
9
9
,96
5
1
200
1E
-
04
9
9
,98
1
3
9
9
,98
1
3
9
9
,98
0
4
9
9
,98
1
3
9
9
,97
8
7
9
9
,97
1
9
9
9
,95
1
5
9
9
,91
3
2
200
1E
-
05
9
9
,56
2
0
9
9
,36
2
9
9
9
,36
8
9
9
9
,39
6
1
9
9
,41
4
8
9
9
,43
1
8
9
9
,48
8
8
9
9
,45
4
0
200
1E
-
06
9
9
,25
8
3
9
9
,26
2
6
9
9
,26
6
0
9
9
,27
5
3
9
9
,26
7
7
9
9
,27
1
9
9
9
,26
9
4
9
9
,26
7
7
200
1E
-
07
9
9
,25
2
4
9
9
,25
3
2
9
9
,25
8
3
9
9
,24
9
0
9
9
,24
9
8
9
9
,25
2
4
9
9
,25
0
7
9
9
,25
1
5
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
Author i
den
ti
fi
cation i
n bi
blio
gr
ap
hic
da
t
a
usi
ng
dee
p ne
ural netw
or
ks
(
Fi
rd
aus
)
917
The
best
res
ults
of
D
NN
s
a
re
a
m
odel
with
on
e
hidden
la
ye
r
DNNs
str
uc
tures
an
d
20
0
neur
on
s
f
or
each
la
yer
with
0.5
dro
pout
va
lue.
F
r
om
t
he
c
la
ssific
at
ion
process,
the
D
N
Ns
model
str
uc
ture
is
sel
ect
ed
base
d
on
the
highest
accurac
y.
Bot
h
trai
ni
ng
da
n
te
sti
ng
proces
se
s
,
bu
t
it
more
imp
or
ta
nt
in
th
e
te
sti
ng
pr
oce
ss.
T
he
highest
a
ver
a
ge
accu
racy
f
or
al
l
data
,
a
bout
99.99
%
in
trai
ning
a
nd
99.98
%
in
te
sti
ng.
T
he
sa
me
resu
lt
s
f
or
al
l
AND
pro
blems
,
the
acc
uracy
va
lue
ab
out
99
%
.
H
oweve
r,
t
he
recall
value
f
or
hom
onym
-
s
ynonym
is
un
der
70%
(see Ta
ble
6
an
d Fi
gure
4).
T
he
num
ber
of sa
mp
le
data
with
a
homon
ym
-
s
ynonym
c
ondit
ion i
s less
tha
n ot
her
conditi
ons
ar
ound 1
10
data from
a
total
of
4419 d
at
a
or
only
ab
out
2.5% of
the
total
data.
The
imbala
nc
e
data
can
decr
eas
e th
e M
L
p
e
rfo
rma
nce.
Table
6
. Pr
opose
d DNNs
classi
ficat
ion
perf
orma
nces
AND P
rob
lem
Accuracy
Precisio
n
Recall
F1
Score
All
9
9
,99
5
1
9
9
,32
8
9
9
9
,32
9
3
9
9
,31
2
3
9
9
,98
3
8
9
7
,98
8
7
9
7
,86
4
1
9
9
,99
1
9
Ho
m
o
n
y
m
9
9
,27
3
0
8
7
,92
8
2
8
8
,96
0
0
8
8
,31
6
3
9
8
,03
0
3
7
9
,83
3
3
7
7
,45
2
4
9
8
,91
9
8
Sy
n
o
n
y
m
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
9
9
,93
8
0
9
3
,37
1
2
9
5
,34
6
3
9
9
,96
8
6
Ho
m
o
n
y
m
-
Sy
n
o
n
y
m
9
9
,58
4
8
8
4
,11
7
6
8
8
,23
5
3
8
5
,62
0
9
9
8
,15
3
8
6
9
,23
0
8
6
9
,23
0
8
9
9
,07
6
9
No
n
-
Ho
m
o
n
y
m
-
S
y
n
o
n
y
m
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
No
te :
■
Tr
ain
in
g
a
n
d
■
Testin
g
Figure
4. The
DNNs
te
sti
ng
Fo
r
validat
in
g
the
propose
d
D
NN
s
a
ppr
oach,
three
te
ch
niqu
es
li
ke
Naïve
B
ayes,
Ra
nd
om
Fo
r
est
, a
nd
su
pp
or
t
vect
or
mac
hin
e
(SV
M
)
are
c
ompar
ed
i
n
te
r
ms
of
accurac
y,
pr
eci
sion,
recall
,
a
nd
F1
-
Sc
or
e
.
Ta
ble
7
sh
ows
the
cl
as
sific
at
ion
s
acc
ur
ac
y
pe
rform
ances
of
t
he
D
NN
s
cl
assifi
e
r
com
par
e
d
with
ot
her
M
L
te
c
hniq
ues,
su
c
h as
99.
8%
for a
ll
ty
pe
s
of
data,
98.
0%
for
homon
ym
,
99.8
%
f
or
s
ynonym,
98.
1%
for
hom
onym
-
s
ynonym
,
and
100%
for
non
-
hom
onym
-
syn
onym.
I
n
al
l
metri
cs
of
pr
eci
sio
n,
reca
ll
,
an
d
F
1
-
sc
or
e,
DNNs
ou
t
pe
rforms
oth
e
r
ML
te
ch
nique.
Actuall
y,
S
V
M
s
deliv
er
a
uniq
ue
s
ol
ution
in
the
cl
assifi
ca
ti
on
ta
s
k
,
since
th
e
opti
mali
ty
pro
blem
is
co
nvex
.
T
his
is
an
adv
a
nta
ge
co
mp
a
red
t
o
A
N
Ns,
w
hich
have
mu
lt
iple
s
olut
ion
s
ass
ociat
ed
wit
h
local
minima
.
H
ow
e
ve
r,
D
NN
s
us
ed
a
de
ep
st
ru
ct
ur
e
of
hi
dd
e
n
la
ye
rs
.
I
t
ca
n
ov
ercome
th
e
drawb
ac
k.
Ther
e
f
or
e,
all
perform
a
nces a
re im
prov
e
d
by
ab
ou
t
1%
over
the SV
M
.
Our
pro
po
se
d
method
with
D
NN
s
cl
assifi
er
in
aut
hor
id
ent
ific
at
ion
on
bi
bliogra
ph
ic
dat
a
co
ntainin
g
homon
ym
a
nd
sy
no
nym
data
pro
du
ce
a
good
pe
rformanc
e.
By
ex
plo
ri
ng
the
res
ult,
our
method
with
D
NN
s
cl
assifi
er
ha
s
a
bette
r
perf
or
m
ance
tha
n
oth
e
r
M
L
te
c
hn
i
qu
es.
Co
mp
a
rin
g
to
the
same
re
search
with
the
sam
e
dataset
[19
]
,
our
pro
po
se
d
D
NN
s
meth
od
ha
s
a
bette
r
res
ul
t
in
a
recall
,
i.e.
,
97.
9%
c
ompare
d
to
Naïve
Ba
yes,
Ra
ndom
Fores
t
cl
assifi
ers,
a
nd
sup
port
ve
c
tor
mac
hi
ne
.
As
s
how
n
in
Table
8
,
the
non
-
s
ynonym
-
hom
onym
cat
egory
wor
ks
perfect
ly
in
al
l
performa
nce
measu
reme
nts,
w
hic
h
has
100%
.
It
is
no
t
s
urpr
isi
ng
becau
s
e
of
the
cat
egory
of
non
-
s
ynon
ym
-
ho
monym
is
not
the
mai
n
issu
e
for
a
utho
r
i
den
ti
ficat
io
n.
We
ex
plaine
d
above,
sy
no
nym
a
nd
homon
ym
a
re
crit
ic
al
prob
le
ms.
T
he
syn
onym
-
hom
onym
cat
eg
ory
ha
rd
e
r
pro
blem
t
o
s
olv
e
.
These i
ss
ues
m
us
t
be
e
xp
l
or
e
d pe
r
eac
h
cat
eg
ory for t
he
a
uth
or i
den
ti
ficat
ion b
y
it
s c
har
a
ct
erist
ic
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
1693
-
6930
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
,
V
ol.
19
,
No.
3
,
June
20
21
:
91
1
-
919
918
Table
7
. C
omp
ariso
n of D
NNs p
e
rformance
s
(
%
)
f
or all
d
at
a w
it
h othe
r M
L tec
hniq
ues
Clas
sifier
Accuracy
Precisio
n
Recall
F1
Score
Naïv
e Bay
es
9
9
,87
1
6
6
7
,92
8
5
7
3
,10
4
7
6
9
,26
7
4
Ran
d
o
m
Forest
9
9
,80
4
4
5
6
,86
9
9
5
8
,23
8
0
5
5
,24
7
0
SVM
9
9
,97
5
3
9
4
,56
7
7
9
5
,23
0
0
9
4
,59
9
7
DNN
9
9
,98
3
8
9
7
,98
8
7
9
7
,86
4
1
9
9
,99
1
9
Table
8
. C
omp
ariso
n of D
NNs p
e
rformance
s
(
%
)
f
or eac
h AND
pro
blems
w
it
h othe
r ML
techn
i
qu
e
s
AND
P
rob
lem
Clas
sifier
Accuracy
Precisio
n
Recall
F1
Score
Ho
m
o
n
y
m
Naïv
e Bay
es
9
7
,20
2
8
3
7
,94
8
7
4
3
,46
1
5
3
9
,64
1
0
Ran
d
o
m
Forest
9
7
,75
5
3
4
2
,46
9
1
4
8
,73
0
2
4
4
,01
1
6
SVM
9
8
,33
3
3
8
1
,83
3
3
8
0
,83
3
3
8
1
,27
7
8
DNN
9
8
,03
0
3
7
9
,83
3
3
7
7
,45
2
4
9
8
,91
9
8
Sy
n
o
n
y
m
Naïv
e Bay
es
9
9
,50
5
9
7
0
,03
6
2
7
9
,35
3
6
7
3
,33
3
8
Ran
d
o
m
Forest
9
9
,43
5
7
4
7
,88
5
9
5
5
,93
4
7
5
0
,33
8
4
SVM
9
9
,89
1
8
9
4
,24
6
0
9
5
,39
5
3
9
4
,12
2
8
DNN
9
9
,93
8
0
9
3
,37
1
2
9
5
,34
6
3
9
9
,96
8
6
Ho
m
o
n
y
m
-
Sy
n
o
n
y
m
Naïv
e Bay
es
9
6
,57
1
4
5
5
,35
7
1
5
7
,14
2
9
5
6
,12
2
4
Ran
d
o
m
Forest
9
4
,90
9
1
2
3
,63
6
4
3
1
,81
8
2
2
5
,75
7
6
SVM
9
8
,00
0
0
7
5
,00
0
0
7
2
,91
6
7
7
3
,80
9
5
DNN
9
8
,15
3
8
6
9
,23
0
8
6
9
,23
0
8
9
9
,07
6
9
No
n
-
Ho
m
o
n
y
m
-
S
y
n
o
n
y
m
Naïv
e Bay
es
9
9
,83
5
7
6
5
,44
4
3
7
0
,33
2
3
6
6
,98
6
3
Ran
d
o
m
Forest
9
9
,75
1
8
4
8
,65
9
3
5
1
,22
0
5
4
8
,39
8
7
SVM
9
9
,98
2
7
9
4
,05
9
4
9
5
,04
9
5
9
4
,38
9
4
DNN
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
1
0
0
,0000
4.
CONCL
US
I
O
N
Fr
om
the
e
xperiment
re
su
lt
s
,
it
can
be
co
nclu
ded
t
hat
the
meth
od
pr
oduces
good
r
esults
for
al
l
pro
blems
with
an
ave
ra
ge
acc
ur
ac
y
of
99.98
%.
The
meth
od
so
lves
t
he
syn
onym
pro
ble
m
bette
r
tha
n
ho
monym;
besides
,
t
he
pe
rformance
r
eg
ard
i
ng
the
c
ombine
d
syn
onym
-
homon
ym
pro
blem
is
sti
ll
le
ss
tha
n
sat
isf
act
ory.
The
c
omplexit
y
of r
ec
ognizin
g
an
d
assi
gn
i
ng pu
blica
ti
on
s
to the r
es
pecti
ve
au
th
or
s is
not a simple t
ask.
So
m
e
te
chn
iq
ues
ha
ve
bee
n
pro
pose
d
f
or
s
olv
i
ng
a
utho
r
na
me
disam
bigu
at
ion
,
s
pecific
al
ly
in
syn
onym
a
nd
homon
ym
pro
blems.
Four
m
achine
le
ar
ni
ng
al
gorith
ms
ha
ve
be
en
c
ompare
d
to
obta
in
preci
se
performa
nce.
The
res
ults
r
ev
eal
ed
that
N
Ns
with
on
e
la
ye
r
si
gn
ific
a
ntly
ou
t
performe
d
oth
e
r
machi
ne
le
arn
in
g
te
ch
niq
ue
s
with
a
n
a
ver
a
ge
acc
ur
ac
y
of
99.
98%.
Sett
ing
up
a
N
Ns
al
go
rithm
is
mu
c
h
more
te
dious
than
us
in
g
an
off
-
the
-
sh
el
f
cl
assi
fier
li
ke
S
VM.
F
or
la
rg
e
-
data
a
nalytic
al
meth
od
s
ass
ociat
ed
with
machi
ne
le
arn
i
ng
al
gor
it
hm
s,
deep
e
r
NN
s
usi
ng
DNNs
a
re
promisi
ng
al
gorit
hms
in
va
riou
s
fiel
ds
of
a
pp
li
cat
io
n,
incl
ud
i
ng
aut
hor
na
me
disam
biguati
on.
DNNs
e
mpl
oy
var
i
ous
de
ep
le
ar
ning
al
gorithms
base
d
on
netw
ork
struct
ur
e,
act
ivati
on
functi
on,
a
nd
model
par
amet
ers,
with
their
outp
ut
dep
e
ndin
g
on
the
da
ta
represe
ntati
on
f
ormat
.
F
r
om
t
he
exp
e
rime
ntal
r
esults
of
this
r
esearch
,
howe
ver,
both
D
NNs
an
d
S
V
M
ob
ta
in
hi
gh
e
r
pe
r
forma
nce
i
n
synon
ym
pro
blems
tha
n
i
n
homon
ym
prob
le
m
s.
With
t
he
propose
d
D
NN
s
,
the
pe
rfo
r
mances
in
s
ynonyms
resu
lt
in
values
for
acc
uracy
,
preci
sion,
recall
,
a
nd
F
1
-
sc
ore
of
99.
94%,
93.
37%,
95.
35%,
and
99.97%,
re
sp
ect
ively
.
In
f
uture
work,
in
the
bi
g
data
era
for
t
he
m
od
e
rn
dig
i
ta
l
li
br
ary
,
D
N
Ns,
the
pro
pos
ed
meth
od
is
typ
ic
al
ly
very
he
lpf
ul
for
wor
king
w
it
h
la
rg
e
r
data
set
s.
Be
sides
,
for
homon
ym
and
hom
onym
-
syn
onym,
a
n
appr
opriat
e
m
et
hod
sh
oul
d
be
im
pl
emented
in
othe
r
dataset
s
a
nd
increa
sed
pe
rformance
.
T
he
us
e
of
featur
e
eng
i
neer
i
ng
ba
sed
on
semanti
c a
ppr
oa
ched f
or ti
tl
e att
ribu
te
c
ou
l
d i
mp
r
ove the
p
e
rformance
of a
ll
cases.
ACKN
OWLE
DGE
MENTS
We
t
hank
the
M
i
nistry
of
Re
search
,
Tec
hnolog
y,
an
d
Higher
E
du
cat
ion
,
Re
pu
blic
of
I
ndonesi
a
(K
e
men
riste
kd
ikti
RI)
,
for
f
un
ding
the
resear
ch
on
“P
eneli
ti
an
Dise
rtasi
D
ok
t
or
”
Re
sea
rc
h
G
ra
nt
with
c
on
t
ract
numb
e
r 2
11
/S
P2H/LT/
DRP
M
/I
V/2019.
REFERE
NCE
S
[1]
A.
A.
Ferr
ei
ra
a
nd
M.
A.
Gonça
lve
s,
“A
Brie
f
S
urve
y
of
Auto
m
at
i
c
Me
thods
for
Author
Nam
e
Disambi
guation,
”
ACM
SIGMO
D
Re
cord
,
vol
.
41
,
no.
2
,
pp
.
15
-
26
,
2012.
[2]
I.
Hus
sain
and
S.
As
ghar
,
“A
survey
of
aut
hor
n
a
me
disam
b
iguati
on
t
ec
hniqu
es:
2
010
-
2016,
”
Kno
wl.
Eng
.
Rev
Th
e
Knowle
dge
Enginee
ring
Revie
w
,
vol.
00
:0, pp. 1
-
24,
2017
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
O
M
N
IKA
Tel
ec
om
m
un
C
ompu
t El
C
on
t
ro
l
Author i
den
ti
fi
cation i
n bi
blio
gr
ap
hic
da
t
a
usi
ng
dee
p ne
ural netw
or
ks
(
Fi
rd
aus
)
919
[3]
S.
Miloj
evi
ć
,
“Ac
cur
a
cy
of
simp
le
,
initials
-
base
d
methods
for
author
name
d
isambiguat
ion
,
”
J
.
In
formetr.
,
vol.
7,
no.
4
,
pp
.
767
-
7
73,
2013
.
[4]
H.
N.
Tra
n
,
T.
H
uynh,
and
T
.
Do,
“Author
n
ame
d
isam
biguation
b
y
using
de
ep
neu
ral
ne
twork,
”
in
Asian
Con
fe
renc
e
on
Intelli
g
ent Inf
orm
ati
on
and
D
atabase
Syst
ems
,
2014,
pp.
123
-
1
32.
[5]
F.
Momen
i
and
P
.
Mayr
,
“Using
C
o
-
aut
horship
Net
works
for
Autho
r
Nam
e
Disa
mbiguati
on
,
”
Proc
.
1
6th
ACM
/IEEE
-
CS
Jt.
Con
f. Dig
it
.
Libr
.
-
JCDL
’16
,
2016
,
pp
.
2
61
-
262
.
[6]
D.
Shin,
T
.
Ki
m
,
J.
Choi
,
and
J
.
Kim,
“Author
n
am
e
disa
mbi
gua
ti
on
using
a
gra
ph
model
wi
th
n
ode
spli
tt
ing
and
me
rging
base
d
o
n
bibl
iog
rap
hi
c i
nforma
t
ion,
”
Sc
i
ent
ometri
cs
,
vo
l. 100, no. 1, pp.
15
-
50,
2014
.
[7]
N.
R.
Smal
h
ei
se
r
and
V.
I.
Torv
ik,
“Author
name
disa
mbi
gua
ti
o
n,
”
Annu.
R
ev.
I
nf.
S
ci.
Te
chnol.
,
vol
.
43,
no.
1,
pp.
1
-
43
,
2009
.
[8]
X.
Li
n
,
J.
Zhu,
Y.
Ta
ng
,
F.
Yan
g,
B.
Peng,
and
W.
Li
,
“A
nove
l
appr
oa
ch
for
au
thor
name
disa
mbi
guation
usi
ng
ran
king
conf
ide
n
ce
,
”
Int
ernati
on
al Confe
ren
ce
o
n
Database
Syst
ems
for
Adv
anc
e
d
Appl
i
cat
ions
,
2
017
,
pp.
169
-
182
.
[9]
I.
Hus
sain
and
S.
As
ghar
,
“L
U
CID:
Author
n
a
me
d
isam
bigu
ation
using
gr
aph
Struct
ur
al
Clus
te
ring
,
”
in
20
17
Inte
lligen
t
Syst
e
ms
Confe
renc
e
,
I
nte
lliSys 2017
,
2
018,
vol
.
2018
-
J
anua
,
pp.
406
-
41
3.
[10]
R.
G.
Cot
a,
A.
A
.
Ferre
ira,
C
.
Nasci
m
ent
o,
M
.
A.
Gonça
lve
s,
and
A.
H.
F.
L
ae
nder
,
“An
unsupervis
ed
heur
ist
ic
-
b
ase
d
hie
rar
chi
c
al
m
ethod
for
n
am
e
di
sambi
guation
in
bibliographic
citat
ions
,
”
J.
Am
.
Soc
.
Inf.
Sc
i.
Te
chnol
.
,
vo
l.
61,
no.
9
,
pp
.
1853
-
1870,
2010
.
[11]
A.
Culot
ta,
P
.
K
ana
ni
,
R
.
Hall,
M.
W
ic
k
,
and
A
.
McC
al
lu
m,
“A
uthor
disa
mbi
gu
at
ion
using
err
or
-
drive
n
machine
le
arn
ing
wi
th
a
ran
king
loss
fun
ct
ion
,
”
in
Sixth
Inte
rnational
W
orkshop
on
Info
rm
at
ion
Inte
grat
ion
on
the
W
e
b
(IIWe
b
-
07), Van
couv
er,
Canada
,
2007.
[12]
A.
A.
Ferr
ei
ra
,
A.
Veloso
,
M
.
A.
Gon
ça
lv
es,
and
A.
H.
F.
Laende
r
,
“
Eff
e
ct
iv
e
self
-
traini
ng
aut
hor
name
disam
biguation
i
n
schola
r
ly
d
igi
t
al
li
bra
rie
s,
”
in
P
roce
edi
ngs
o
f
the
10th
annual
jo
in
t
con
fe
ren
ce
on
Digit
al
li
brar
ie
s
,
2010,
pp
.
39
-
48
.
[13]
J.
Ta
ng
,
A.
C.
M.
Fong,
B.
W
a
ng,
and
J.
Zha
n
g,
“A
un
ifi
ed
pr
obabi
li
st
ic
fra
m
ework
for
na
me
disam
bigu
ation
in
digi
tal
li
bra
ry
,
”
I
EE
E
Tr
ans.
Knowl.
Data
Eng
.
,
v
ol.
24
,
no
.
6
,
pp
.
975
-
987,
2011
.
[14]
H.
Han
,
L
.
Gil
es,
H.
Zha,
C.
L
i,
and
K.
Ts
iout
siouli
k
li
s,
“
Two
supervis
ed
l
ea
rn
ing
appr
o
ac
hes
for
na
me
disam
biguation
i
n
au
thor
ci
t
at
ion
s,”
in
Proceedi
n
gs
of
the
2004
J
oint
ACM/
I
EE
E
Confe
renc
e
on
Digit
al
Libr
arie
s,
2004.
,
2004,
pp.
296
-
305.
[15]
S.
F.
Schifa
no
,
T.
Sgarba
n
ti,
an
d
L.
To
ma
ss
etti,
“Author
ship
re
cogni
ti
on
and
di
sambi
guation
of
scie
nt
ifi
c
p
ape
r
s
using a
n
eur
a
l
n
e
tworks a
pproa
ch
,
” in
Proceedi
ng
s of
Sc
ie
n
ce
,
201
8.
[16]
S.
S.
Khan
and
M.
G.
Madde
n
,
“A
survey
of
author
na
me
disambiguat
ion
t
ec
hn
i
ques:
2010
–
201
6,
”
The
Know
le
d
ge
Engi
ne
ering
R
evie
w,
vol
.
32
,
201
7
.
[17]
H.
Han
,
H.
Zha
,
and
C
.
L.
Gil
es,
“Na
me
disa
mbi
g
uat
ion
spec
tra
l
in
aut
ho
r
ci
t
at
ions
using
a
k
-
way
clus
te
ring
method,
”
in
Proc
ee
dings
o
f
th
e
ACM
/IEEE Joi
nt
Confe
ren
c
e
on
Dig
it
al
Lib
rar
ie
s,
JCDL,
De
nve
r,
CO
,
USA
,
2005,
pp
.
7
-
11
.
[18]
H.
Han,
W.
Xu,
H.
Zh
a,
and
C
.
L.
Gi
le
s,
“A
hi
e
rar
chica
l
na
ive
Baye
s
mi
x
ture
mode
l
fo
r
na
me
disam
biguation
i
n
aut
hor citati
ons
,
”
in
Proc
ee
dings
of the
2005
AC
M
sympos
ium
on
Applied comput
ing
,
2005
,
pp
.
10
6
5
–
1069.
[19]
J.
Kim
and
J.
Ki
m,
“T
h
e
im
pa
ct
of
im
b
al
an
ce
d
tr
ai
ning
data
on
m
ac
hin
e
le
arn
ing
f
or
aut
hor
na
me
disam
biguation
,
”
Sci
en
tomet
rics
,
vol.
117
,
no
.
1
,
p
p.
511
-
526
,
201
8.
[20]
P.
Mitr
a,
J.
Kang
,
D.
Le
e
,
and
B.
On,
“Co
mpa
r
at
i
ve
study
of
name
disam
b
igu
a
ti
on
proble
m
using
a
sca
la
b
le
bloc
k
in
g
-
base
d
fr
am
ewor
k,
”
in
Proceedi
n
gs
of
th
e
5
th
AC
M/IE
E
E
-
CS
Joi
n
t
Conf
ere
nce
on
Digit
al
Libr
aries
(JCDL’05)
,
200
5,
pp.
344
-
353
.
[21]
T.
Bac
k
es,
“T
h
e
im
pa
ct
of
n
am
e
-
ma
t
chi
ng
and
bl
ocki
ng
on
aut
hor
disam
b
igua
t
ion,”
in
Pro
c
ee
d
ings
of
the
27th
AC
M
Inte
rnational
Co
nfe
renc
e
on
Info
rm
ati
on
and
Kn
owle
dge
Manag
eme
nt
,
2018,
pp.
803
-
812.
[22]
Q.
Qin,
J.
L
i,
L
.
Zh
ang,
Y.
Yue,
and
C.
Li
u,
“C
ombi
ning
Low
-
d
im
ensional
Wa
v
el
e
t
Fe
at
ur
es
an
d
Support
Vec
to
r
Mac
hine for
Arr
hythm
ia Beat
Cl
assi
fic
a
ti
on,
”
S
ci.
R
ep.
,
vo
l. 7, no
.
1
,
pp
.
1
-
12
,
20
17.
[23]
D.
Han,
S.
L
iu,
Y.
Hu,
B.
W
ang,
and
Y.
Sun,
“ELM
-
base
d
n
am
e
disam
biguation
i
n
bibliography
,
”
World
Wide
W
eb
,
vol.
18
,
no
.
2
,
pp
.
253
-
263
,
2015
.
[24]
W.
Li
u
et
al
.
,
“
A
survey
of
d
eep
neur
a
l
n
et
wor
k
arc
h
itect
ur
es
a
nd
their
appl
i
ca
t
ions,”
N
euroc
o
mputing
,
vo
l.
23
4,
pp.
11
-
26
,
2017
.
[25]
M.
Sokolova
and
G.
La
pa
lme,
“A
sys
te
ma
tic
analy
sis
of
per
for
ma
n
ce
measures
for
cl
assifi
ca
t
ion
ta
s
ks,”
Inf
.
Proce
ss
.
Manag.
,
vol. 45, no. 4, pp. 427
-
4
37,
2009
.
Evaluation Warning : The document was created with Spire.PDF for Python.