TELKOM
NIKA
, Vol.11, No
.3, March 2
0
1
3
, pp. 1213 ~ 1220
ISSN: 2302-4
046
1213
Re
cei
v
ed O
c
t
ober 5, 20
12;
Revi
se
d Ja
n
uary 5, 2013;
Acce
pt
ed Jan
uary 18, 201
3
Information Extraction from Research Papers based o
n
Conditional Random Field Model
Zhu Shuxin*,
Xie Zhonghong, Chen Y
u
ehong
Coll
eg
e of Information Sci
enc
e and T
e
chno
l
o
g
y
, Nan
jin
g A
g
ricult
ural U
n
iv
ersit
y
W
e
iga
ng 1, Na
njin
g, Chi
na, 2
100
95
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zsx@n
ja
u.ed
u.cn
A
b
st
r
a
ct
With the increasing us
e of Cit
eSeer
academ
ic s
earch engines, the accur
a
cy of such system
s
has
beco
m
e mor
e
and more i
m
p
o
rtant. T
he paper ado
pts t
he improv
ed parti
cle sw
arm opti
m
i
z
at
io
n algor
it
h
m
for trainin
g
co
nditi
ona
l ran
d
o
m fie
l
d
mo
de
l and a
p
p
lies
it into the res
earch p
a
p
e
rs
’ title and citat
i
on
retrieval. T
he i
m
pr
ove
d
parti
cl sw
arm opti
m
i
z
at
io
n
alg
o
ri
thm bri
ngs the
particle sw
arm ag
gre
gatio
n
to
preve
n
t particl
e sw
arm from
bei
ng pl
un
ge
d
into loca
l
con
v
erge
nce too
early, an
d use
s
the line
a
r in
erti
a
factor and le
ar
nin
g
factor to update p
a
rticle r
a
te. It can c
ontrol alg
o
rith
m in
infinite
iter
atio
n by
the iterati
o
n
betw
een p
a
rtic
le rel
a
tive p
o
si
tion cha
n
g
e
ra
te. T
he re
sults
of w
h
ich usin
g the
stand
ard
research
pap
ers
’
hea
ds a
nd r
e
ferenc
es to ev
a
l
uate th
e trai
n
ed co
nd
itio
n
a
l
rand
o
m
fiel
d
mo
de
l show
s t
hat co
mp
are
d
w
i
th
traditio
nal
ly co
nditi
ona
l rand
o
m
fiel
d mode
l and Hi
dd
en M
a
rkov Mod
e
l, the con
d
itio
na
l rand
o
m
field
mode
l,
opti
m
i
z
e
d
an
d traine
d by impr
oved p
a
rticle s
w
arm, has be
e
n
better amel
io
rated in the as
pect of F
1
me
an
error and word error rate.
Key W
o
rd
s:
Con
d
itio
nal
R
and
o
m
F
i
el
ds
Model; Partic
le Sw
arm Opt
i
mi
z
a
tio
n
; Max
i
mu
m Lik
e
li
ho
od
Param
e
ter Estim
a
tion; In
for
m
ation Extractio
n
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Acade
mic
se
arch en
gine
s such a
s
Cite
Seer an
d Co
ra ha
s b
r
oug
ht great conv
enien
ce
and influen
ce
for rese
arch
ers. So the q
uality prov
ide
d
by such
system
s is very important. It
is
depe
ndant o
n
the informa
t
ion retrieval
comp
one
nts
whi
c
h extra
c
t
titles, authors, institution
s
and
other meta-d
ata from pap
ers be
ca
use
these me
tadata later would be use
d
for a variety o
f
appli
c
ation
s
, su
ch a
s
dom
ain ba
sed
se
arching a
nd the autho
rs
re
sea
r
ch.
Previou
s
ly, acad
emic info
rmation
retri
e
val is
ba
se
d on t
w
o m
a
in ma
chi
n
e
learning
techni
que
s. The first is Hi
d
den Ma
rkov Model
(HMM
). HMM obtain
s
gen
eratio
n model from t
h
e
input and lab
e
l sequ
en
ce pairs. Although HMM
gains gre
a
t su
ccess,
the standard HMM mode
l
is difficult to build the mo
del for multip
le indep
end
e
n
t feature
s
of obse
r
vation
seq
uen
ce. T
he
other te
chni
q
ue is b
a
sed
on su
ppo
rt vector m
a
ch
in
e (SVM) cl
assifier. It can
pro
c
e
ss m
u
ltiple
non-i
nde
pen
d
ent feature
s
,
but it divides i
n
formatio
n re
trieval into two step
s a
nd b
r
ea
ks th
e cl
o
s
e
coo
r
din
a
tion
betwe
en stat
e transitio
n a
nd ob
servatio
n.
This p
ape
r i
n
trodu
ce
s a
con
d
itional
random
fiel
d
model [1] (CRF) fo
r the
aca
demi
c
metadata retrieval and app
lies the mode
l to solving
th
e practi
cal problem
s in ord
e
r to prove that
the model i
s
sup
e
rio
r
to HMM and SV
M. CRF h
a
s
been
su
ccessfully use
d
in
biologi
cal en
tity
recognitio
n
[2, 3, 4]. The
model loo
s
en
s the
stron
g
random hypot
hesi
s
of HM
M and effectively
overcome
s th
e label bi
as
p
r
oble
m
. Similar to Max Ent
r
opy Ma
rkov Model
(MEM
M). CRF is
al
so
a co
nditional
prob
ability se
quen
ce m
ode
l. But the di
fferen
ce i
s
that
CRF i
s
a
n
u
ndire
cted
gra
p
h
model.
In CRF, m
a
ximum likeliho
od pa
ramete
r estimation i
s
cru
c
ial
be
ca
use th
ese pa
ramete
rs
con
c
e
r
n the
perfo
rman
ce
and a
c
cura
cy of the ap
plicatio
ns b
a
sed o
n
CRF
.
The maxim
u
m
likeliho
od est
i
mation usu
a
l
l
y adopts no
nlinea
r conj
u
gate gradi
ent
algorithm [5, 6, 7], Newton
method [8], BFGS [9], g
r
adie
n
t accel
e
ration al
gor
ithm [11], virtual eviden
ce accele
ratio
n
algorith
m
[1], piece
w
ise hypothe
sis likeli
hood me
tho
d
[10, 11], stocha
stic
g
r
adi
ent method [12],
minimum div
e
rge
n
ce bea
m method [13
]
and so on.
This pa
per
use
s
improved parti
cle swarm
optim
ization alg
o
ri
thm to estimate the
maximum likelihoo
d para
m
eters of CRF model. In o
r
de
r to preve
n
t the algorithm fallen into the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1213 – 1
220
1214
local co
nverg
ence in the
early time of s
earch, this paper adopt
s an appro
p
riate aggre
gat
ion
degree to control particl
es aggreg
ation
level. A
nd to prevent the algorit
hm convergi
ng slo
w
ly
near the b
e
st location
an
d goin
g
into
an infinite it
eration, thi
s
pape
r em
plo
y
s ine
r
tia a
nd
learni
ng facto
r
whi
c
h is lin
ear varying t
o
cont
rol part
i
cle se
arch range an
d the relative cha
nge
rate of iterati
v
e particle p
o
sition to terminate it
erati
on. Finally, this arti
cle ap
plies the trained
con
d
itional
ra
ndom fiel
d m
odel i
s
the
use of t
he
stan
dard
re
se
arch pap
ers’ h
e
ad an
d refere
nce
data set retri
e
val. The experim
ental re
sults
sho
w
s that comp
are
d
with traditi
onally con
d
itional
rand
om field model an
d Hi
dden Ma
rkov Model, the
condition
al ran
dom field mo
del ,optimize
d
and traine
d by improved particl
e swarm, has been
better ameliorated in the
aspe
ct of w
o
rd
ac
cur
a
cy
an
d
F met
r
ic
s.
This p
ape
r i
s
orga
nized a
s
follo
ws: Se
ction 1 i
n
trod
uce
s
the
co
n
c
ept of
CRF
and its’
maximum likelihoo
d para
m
eter estima
tion, Sect
ion 2 descri
bes the improved particle swarm
optimizatio
n and p
r
opo
se
s a ne
w mod
e
l para
m
eter
s estim
a
tion
algorith
m
of CRF
ba
sed o
n
it,
Section 3 sho
w
s the exp
e
ri
mental re
sult
s and Se
ction
4 gives the summari
zatio
n
of this paper.
2. Condition
al Random F
i
eld Model
Con
d
itional random field model (CRF
) is
an undirected graph mo
del used to calcul
ated
the condition
al proba
bility of
the output s
eque
nce
under the premi
s
e of the given input
seq
uen
ce. Lafferty [1] gives the definition of
CRF: For a given observation seque
nce x, the
con
d
itional probability of its’ tag sequ
e
n
ce y is
the
norm
a
lized p
r
odu
ct of the bit
function, as
sho
w
n in Fo
rmula (1
).
(
|
)
e
xp(
(
,
,
,
)
(
,
,
)
)
1
p
y
x
t
y
y
xi
s
y
xi
jj
i
i
k
k
i
j
k
(1)
whe
r
e
(,
,
,
)
1
ty
y
x
i
ji
i
is the tran
sition
feature fun
c
t
i
on of the whol
e ob
servation
seq
uen
ce,
y
i
is the label of t
he tag sequ
e
n
ce I,
(,
,
)
s
yx
i
ki
is the
state feature
function
of i , x is
the ob
se
rvation sequ
en
ce
，
j
and
k
are
resp
ectively the weight p
a
ramete
r relat
ed to the
transitio
n feature fun
c
tion a
nd state featu
r
e functio
n
.
Comm
on u
n
d
ire
c
ted g
r
a
p
h
model i
s
a
linear
ch
ain
and
co
rre
sp
ond
s to a fini
te state
machi
ne whi
c
h
is suitabl
e
for
tag se
que
nce. We
sim
p
ly the tran
sit
i
on feature fu
nction
and
st
ate
feature fun
c
tion as b
e
lo
w:
(,
,
)
(
,
,
,
)
1
s
yx
i
s
y
y
x
i
ii
i
(,
)
(
,
,
,
)
1
1
n
Fy
x
f
y
y
x
i
ji
i
j
i
(2)
whe
r
e
(,
,
,
)
1
f
yy
x
i
ji
i
could
be taken
as state fun
c
tion
(,
,
,
)
1
s
yy
x
i
ii
or tran
sition
function
(,
,
,
)
1
ty
y
x
i
ji
i
.Thus, for an observation seq
u
e
n
ce X,
the probability of its’ tag sequ
en
ce
y is expresse
d as follo
ws:
1
(
|
,
)
exp(
(
,
)
)
()
py
x
F
y
x
jj
Zx
j
(3)
Among them,
Z ( X ) is a normali
zatio
n
factor:
()
e
x
p
(
(
,
,
)
)
,
Z
xf
c
y
x
kk
c
xy
k
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Inform
ation Extra
c
tion from
Research Pa
pers ba
sed o
n
Con
d
itional
Ran
dom
… (Zhu Shu
x
in)
1215
For the maxi
mum likelih
o
od paramete
r
estimation of
the condition
al rand
om field model
as
sho
w
n i
n
Formul
a(2), it
is ju
st a p
r
o
c
ess by trai
nin
g
data
s
et
(
1
)
(
1
)
()
()
{(
,
)
,
.
..,
(
,
)
}
nn
Dy
x
y
x
to
estimate th
e
parameter
(,
,
.
.
.
)
12
. Among the
m
, D is
gen
erated by em
p
i
rical
di
stribut
ion
(,
)
px
y
in orde
r to maximize the lo
g likelih
ood of
the training d
a
ta.
3. Conditio
n
al Rand
om Field Model
Parame
ter Estimation based on
Improv
ed
Pa
rticle
S
w
arm Opt
i
miz
a
t
i
on
3.1. Impro
v
e
d
Particle Sw
arm Optimi
z
a
tion Algori
t
hm
Particle
swarm optimizatio
n algo
rithm [
14
], first de
si
gned
by Ken
nedy an
d Eb
erha
rt,
imitates bird
s’ prey behavi
o
r to
solve the optimization pro
b
lem.
Particle swa
r
m optimizatio
n
algorith
m
ma
ke
s a pa
rticle
swa
r
m (
i
x
) se
arch the opti
m
al value a
c
cordin
g to a ce
rtain evolutio
n
rule by initializing it. In each roun
d of evolut
ion, each particl
e'
s velocity and posit
ion have bee
n
update
d
in accord
an
ce wi
th their cu
rre
nt veloci
ty and position, lo
cal an
d glob
al optima value,
s
h
ow
n
as
fo
llo
w
s
:
(1
)
(
)
()
(
)
1
()
(
)
2
vt
w
v
t
id
id
c
r
and
p
x
id
id
c
r
and
p
x
gi
d
(5)
(1
)
(
)
(
1
)
xt
xt
vt
id
id
id
(6)
Among them
，
id
p
,the local optimal value
,
refers to the optimal value got by particle
i
in the dimen
s
ion d hithe
r
to
.
g
p
,the global
optimal value
，
refe
rs to th
e optimal val
ue got by all
particl
es searchin
g the
wh
ol
e search
space up to
n
o
w.
()
id
vt
stands f
o
r particl
e
i’s velocity in
dimen
s
ion d
and
()
id
x
t
rep
r
e
s
ents p
a
rticl
e
i’s current l
o
catio
n
in di
mensi
on d.
(1
)
id
xt
denote
s
pa
rti
c
le i’
s next locatio
n
in di
mensi
on d.
w
is the ine
r
tia
factor a
nd
1
c
is the local
learni
ng facto
r
.
2
c
is the glob
al learni
ng fa
ctor an
d
()
rand
is th
e rand
om fun
c
tion in the ra
nge of
[0,1].
The traditio
n
a
l
particle
swarm optimizatio
n algorith
m
h
a
s the follo
wi
ng pro
b
lem
s
:
1. Premature conve
r
ge
nce to the local
o
p
timal value instea
d of the global
In orde
r to p
r
event the
search p
r
em
aturel
y into lo
cal conve
r
g
e
n
c
e, we use
particl
e
swarm ag
gre
gation [15] to control it. Particle
swa
r
m
aggreg
ation descri
b
e
s
particle disp
ersi
on
defined a
s
:
(
)
m
a
x{|
|
,
,
1
,
2,
..
.,
;
;
1
,
2
,
..
.,
}
dt
x
x
i
j
m
i
jd
N
id
jd
Among them
，
m stand
s for the size of the particl
e swarm and n repre
s
e
n
ts the
search
spa
c
e dime
n
s
ion.
x
id
is denot
es the parti
cl
e i in dimensi
on d and
x
j
d
de
notes the pa
rticle j.
If
()
dt
is less than
the set thre
shold value e,
then
we rei
n
itialize p
a
rti
c
le swa
r
m’
s velocity and
positio
n in d-dimen
s
ion
a
l space.
2. Particle
swarm se
arch
es in
itially ve
ry quickly, ho
wever,
whe
n
clo
s
e to the
optimum
positio
n, its searchin
g spe
ed is slow, a
nd even
in th
e risk of bein
g
trapped into local unlimi
t
ed
iteration.
In the particl
e swarm
optimiz
ation algo
rithm,
param
eter
12
,,
wc
c
is crucial to the speed
of search a
n
d
conve
r
ge
nce. A big inertia factor
do
es good to glob
al sea
r
ch an
d a small ine
r
tia
factor is beneficial to local sear
ch. Its’ e
x
periential va
lue is close
to 1 in
the early search sta
g
e
and clo
s
e to
0.4 in the final stage. Thi
s
paper
ch
o
o
ses the line
a
rl
y decre
asi
ng
inertia facto
r
in
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1213 – 1
220
1216
orde
r to acqu
ire better glo
bal se
arch a
b
ility in
the e
a
rly stage a
n
d better local
search a
b
ility in
the final stag
e.
W (t) =
1- 0.6*t/M
(0
~
1
)
tM
(7)
Whe
r
e M is t
he maximum
numbe
r of iterat
ion
s
and t is the cu
rrent iteration.
Traditio
nal p
a
rticle
swa
r
m
optimizatio
n
algorith
m
s
usu
a
lly set
1
c
and
2
c
as
con
s
tant
numbe
r. In o
r
de
r to ma
ke
the pa
rticle
swarm
co
nve
r
ge
as
so
on
as p
o
ssibl
e
i
n
the e
a
rly la
rge
sea
r
ch spa
c
e
without expl
osive p
r
oliferation and
converg
e
to the optimal v
a
lue a
s
soon
as
possibl
e in the final small
sea
r
ch spa
c
e
without
revol
v
ing aro
und t
he optimal v
a
lue, we
sele
ct
Formul
a (8
) to update
1
c
and
2
c
:
12
2,
2
tt
cc
M
M
(8)
Finally, to en
sure that the
algorith
m
is not
trapped into infinite iteration near the optimal
positio
n, the
pape
r adopt
s
the rate of r
e
lative pos
itio
n chan
ge to termin
ate iteration. When
D is
less than the
threshold val
ue
e, iteration
will be ca
nce
l
ed.
D = (
(1
)
xt
id
-
()
x
t
id
) – (
()
x
t
id
-x
id(t-1
)) / (
()
x
t
id
-x
id(t-1
)
(9)
Among the,
()
x
t
id
stand
s for th
e cu
rrent po
sition of p
a
rt
icle i in di
m
ensi
on d
an
d
(1
)
xt
id
represents t
he next po
sit
i
on of pa
rticl
e
i in
dime
nsion d an
d xid
(
t-1
)
is the
previous
positio
n.
3.2. Conditi
onal Ra
ndo
m Fields Model Param
e
ter Es
timation Meth
od
Bas
e
d on
the
Impro
v
ed Particle S
w
arm
Optimiz
a
tion Algorithm
The pap
er le
ts the para
m
eter vecto
r
of the conditio
nal ran
dom f
i
eld model of
as
particl
e grou
p
and make them in a d
- dimensi
onal
sp
ace search fo
r optimal values for trainin
g
con
d
itional ra
ndom field
s
m
odel. Detail
ed
steps a
r
e de
scribe
d as foll
ows:
First
step: Su
ppo
sing the
current iteratio
n
0,
(
0
)
0
id
tv
and the di
mensi
on n
u
m
ber of
search space is d, we initialize the pa
rticl
e
swarm X co
ntaining m pa
rticle
s:
{,
,
.
.
.
,
}
12
{,
,
.
.
.
,
}
12
(0
~
1
)
x
id
d
d
nd
random
random
random
n
im
We set the local optimal posit
ion and gl
obal optimal positio
n as
：
,0
id
i
d
g
px
p
. Then
we preserve
f
it
ne
ss
id
,the local ada
ptive optimal value, in the array
[]
[
]
Fi
d
and sa
ve the global
adaptive opti
m
al value in the variabl
e
g
F
.
S
e
con
d
st
ep
：
Up
dating th
e local a
nd gl
obal optimal
positio
n
Cal
c
ulating th
e adaptive va
lue of each p
a
rticle:
(,
)
(
,
)
(
)
l
o
g
(
)
f
itne
s
s
p
x
y
F
y
x
p
x
Z
x
jj
id
j
We compa
r
e
the adaptive value of particl
e I in dimensi
on
d with the value of
array
[]
[
]
Fi
d
.If the c
u
rrent value is
greater than
[]
[
]
Fi
d
, then we reset
[]
[
]
Fi
d
as the curre
n
t
value. If the current value i
s
gre
a
ter tha
n
g
F
, then we re
set
g
F
as the cu
rre
nt value.
Third
-
ste
p
: updating the
search velo
city
and po
sition
of the particle
swa
r
m:
Vid(t+1
)
= (1
- 0.6*t/M) * Vid(t) +
(2-t/M)
*
rand
()
*(Pid -
Xid) + (2
+t/M)*ra
nd
()*
(Pg -
Xid)
(1
)
(
)
(
1
)
xt
xt
vt
id
id
id
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Inform
ation Extra
c
tion from
Research Pa
pers ba
sed o
n
Con
d
itional
Ran
dom
… (Zhu Shu
x
in)
1217
Forth
-
setp
：
calcul
ating the
particle
swarm’s agg
re
gati
on
()
dt
:
(
)
m
a
x{|
|,
,
1
,
2
,
.
..,
;
}
22
ma
x
{
(
)
.
.
.
(
)
,
,
1
,
2
,
.
.
.
,
;
}
11
2
m
a
x{
(
)
,
,
1
,
2
,
..
.,
;
}
1
dt
x
x
i
j
m
i
j
id
jd
ij
m
i
j
di
dj
di
n
d
j
n
n
ij
m
i
j
dik
d
jk
k
If
()
dt
is le
ss th
a
n
the p
r
e
s
et t
h
re
shol
d valu
e
e, then
we
reinitialize
the
particl
e
swarm’s
velocity and p
o
sition in the
d-dim
e
n
s
iona
l spa
c
e.
Fifth-s
tep:
1
tt
.
If D< 1
0
-7 or
t = M, then
th
e iteratio
n wil
l
be terminat
ed. Othe
rwi
s
e turn to
the
se
con
d
-
step.
4. Experiment
4.1. Experimental Data
We sel
e
ct
s two re
sea
r
ch pape
rs’ data
sets a
s
our e
x
perime
n
tal data. One is containin
g
the head p
o
rt
ion and the o
t
her is
contai
ning citatio
n
s from referen
c
e
s
. The two
data sets h
a
v
e
been u
s
e
d
for multiple re
se
arch stan
da
rd
test data
（
M
c
Callum et al., 2000; Han e
t
al., 2003.
4.1.1. Data S
e
t of
Rese
ar
ch Paper’s Header
Acade
mic th
esi
s
title is d
e
fined a
s
all
words from
the start of t
he pap
er to
the first
cha
p
ter (usu
ally the intro
ductio
n
) o
r
to
the end of
the first pa
ge.
The title se
ction contai
ns 15
retrieval do
m
a
in: title, author, unit, addre
ss, su
mm
ary, e-mail, date, abstra
c
t, introdu
ction,
telepho
ne nu
mber, key
w
ord, URL, degree, ISSN,
pa
ge numbe
r. The heade
r da
ta set contain
s
935 he
ade
rs.
Acco
rdi
ng to the previo
us
study, we
rand
omly sele
ct 500 pie
c
e
s
of data as traini
ng
data set an
d the remai
n
ing
435 pie
c
e
s
a
s
test data se
t. We call the data set H.
4.1.2. Data S
e
t of
Rese
ar
ch Paper’s Refer
e
nce
s
Referen
c
e
s
data set is g
enerated by Cora
(M
cCall
u
m et al., 2000). It contai
ns 500
referen
c
e ite
m
s an
d we select
s 35
0 items of them a
s
the trai
ning
data set. Th
e rem
a
ining
15
0
items a
r
e tre
a
ted a
s
te
st data. Refe
re
nce
s
contai
n
13 do
main
s:
author, title,
editor, b
o
o
k
title,
date, journal,
volume, sch
ool, institutio
ns, pag
e nu
mber, ad
dress, pre
s
s, sum
m
ary. We
cal
l
the
data set R.
4.2. Validation Techniqu
e
In order to get a comprehe
nsive evaluati
on,
we used
different method
s to comp
are the
test re
sults. I
n
addition to
adoptin
g pre
v
iously
used
word a
c
cura
cy measu
r
em
ent method,
we
will also use a domain F
measurement
method. The
two method
s help ea
ch ot
her a
nd do b
e
tter
to the evaluation re
sults.
1. Word accu
racy. The
nu
mber of the
word
s is
d
e
fine
d as A whi
c
h
belon
g to a p
a
rticul
ar
domain an
d are co
rrectly identified in the domai
n. The numbe
r of the words is defined as B
whi
c
h not belong to but id
entified in a p
a
rticul
ar
dom
ain. The number of the words is defin
ed
as
C whi
c
h belo
ng to a particula
r domai
n
but are not
corre
c
tly iden
tified in the d
o
main and i
s
D
whi
c
h not be
long to and are not identi
f
ied in a do
main. The word accu
ra
cy
is calcul
ated
b
y
. The ratio of the number of
the word
s
who
s
e re
di
ctive identification and re
a
l
identificatio
n are the sam
e
to the total numbe
r of words, is d
e
fined a
s
the comp
re
hen
si
ve
ac
cur
a
cy
.
2. F1mea
s
u
r
ement. We
use th
e pre
c
isi
on rate
and recall rate to cal
c
u
l
ate F1
measurement
. Preci
s
ion
ra
te is exp
r
e
s
sed a
s
and re
call rate
i
s
d
enoted as
.F1 is
r
e
pr
es
e
n
t
ed
a
s
. All domains’ ave
r
ag
e
F1 mea
s
u
r
e
m
ent is
defin
ed a
s
the
averag
e F metric.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1213 – 1
220
1218
4.3. Experiment Result
This p
ape
r li
sts the
com
pari
s
on
re
sul
t
s amon
g th
e co
nditional
rand
om fiel
d model
trained
by the improved
particle swa
r
m o
p
timiza
tion
al
gorithm, HM
M and the
co
nditional
rand
om
field mode trained by the
traditional pa
rticle sw
a
r
m optimizatio
n algorith
m
. Table 1 report
s
the
results ba
se
d
on H data se
t. Similar to the previo
us e
x
perime
n
ts
（
Seymore et a
l
., 1999; Han
et
al., 2003, it e
x
clude
s the introdu
ction fi
eld and pa
g
e
number field
.
F1 measure
m
ent results
is
cal
c
ulate
d
by the precisi
o
n rate and recall rate
of the original experim
ental
report. Tabl
e2
repo
rts the
re
sults b
a
sed o
n
R data set. CRF
（
P
）
d
enote
s
the co
nditional rand
om field mod
e
l
trained by th
e improved
particl
e swa
r
m optimizatio
n algorithm
and CRF
（
T
）
denote
s
the
traditional
co
nditional rand
om field mod
e
l. It c
ould b
e
found from
the compa
r
i
s
on of data , th
e
con
d
itional
ra
ndom field m
odel trai
ned
by the
impro
v
ed parti
cle
swarm o
p
timi
zation al
go
rithm
behave
s
as
good a
s
or b
e
tter than HMM and the tr
aditional co
nditional ra
nd
om field model in
the asp
e
ct of
word a
c
cu
ra
cy, F1 measu
r
eme
n
t. Mo
re
over it is sign
ificant
ly better than the oth
e
r
two model
s in
F measu
r
em
ent and comp
rehe
nsive a
c
curacy.
Table 1. Te
st results ba
se
d
on H data se
t
HMM
CRF
（
P
）
CRF
（
T
）
acc.
F
1
acc.
F
1
acc.
F
1
title
98.2
82.2
99.7 97.1 98.9
96.5
author
98.7
81.0
99.8 97.5 99.3
97.2
unit 98.3
85.1
99.7
97.0
98.1 93.8
address 99.1
84.8
99.7
95.8
99.1 94.7
summar
y
97.8
81.4
98.8
91.2
95.5 81.6
e-mail 99.9
92.5
99.9
95.3
99.6 91.7
date 99.8
80.6
99.9
95.0
99.7 90.2
abstract 97.1
98.0
99.6
99.7
97.5 93.8
Telephone
-numb
e
r
99.8
53.8
99.9
97.9
99.9 92.4
keyword
98.7
40.6
99.7
88.8
99.2 88.5
URL
99.9
68.6
99.9
94.1
99.9 92.4
degree
99.5
68.8
99.8
84.9
99.5 70.1
ISSN
99.8
64.2
99.9 86.6 99.9
89.2
Average
F
75.6
93.9
89.7
Compreh
ensive
accur
a
c
y
93.1%
98.3
%
92.9%
Table 2. Te
st results ba
se
d
on R data se
t
HMM
CRF
（
P
）
CRF
（
T
）
acc.
F
1
acc. F
1
acc.
F
1
author
96.8
92.7
99.9
99.4
98.9 96.5
Book title
94.4
0.85
97.7
93.7
99.3 97.2
date 99.7
96.9
99.8
98.9
98.1 93.8
editor
98.8
70.8
99.5
87.7
99.1 94.7
institution
98.5
72.3
99.7
94.0
95.5 81.6
journal 96.6
67.7
99.1
91.3
99.6 91.7
address 99.1
81.8
99.3
87.2
99.7 90.2
summar
y
99.2
50.9
99.7
80.8
97.5 93.8
Page number
98.1
72.9
99.9
98.6
99.9 92.4
press 99.4
79.2
99.4 76.1
99.2
88.5
college 98.8
74.9
99.4
86.7
99.9 92.4
title
92.2
87.2
98.9
98.3
99.5 70.1
volume 98.6
75.8
99.9
97.8
99.9
89.2
Average
F1
77.6
91.5
89.7
Compreh
ensive
accur
a
c
y
85.1%
95.37
%
92.9%
5. Conclusio
n
The conditio
n
a
l ran
dom fie
l
d model h
a
s great
comp
e
t
itiveness in
POS tagging,
natural
langu
age pro
c
e
ssi
ng and other fields, but its
performance largel
y depends o
n
the param
eter
estimation m
e
thod. A goo
d paramete
r
estimation
al
gorithm i
s
very important for the conditi
onal
rand
om field
model. This paper u
s
e
s
the impr
ove
d
particle
swa
r
m optimization algo
rithm
to
estimate the
maximum likelih
ood pa
rameter of th
e con
d
itional
rando
m field model. An
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Inform
ation Extra
c
tion from
Research Pa
pers ba
sed o
n
Con
d
itional
Ran
dom
… (Zhu Shu
x
in)
1219
comp
ared
with HMM
an
d the tra
d
itional
con
d
it
ional
ra
ndom m
odel,
the conditio
n
a
l ra
ndom
fiel
d
model traine
d by the im
proved
pa
rticle swar
m op
timization al
g
o
rithm h
a
s t
he better
wo
rd
accuracy an
d
F measu
r
em
ent value.
In order to prevent the particle swarm o
p
ti
mization al
gorithm bein
g
fallen into local
conve
r
ge
nce
in early pe
rio
d
, the pape
r
introdu
ce
s th
e parti
cle swarm’
s agg
re
g
a
tion to co
ntro
l
the converge
nce of the algorithm.
To make the pa
rticle swarm restrai
n
from being trap
pe
d into
an infinite ite
r
ation in the optimal positi
on, we
introd
uce an iterati
v
e logarithmi
c
likeliho
od ratio
as the stop
criterion. And to get
the best
search p
e
rfo
r
man
c
e of
the particl
e swarm optimi
z
at
ion,
we ad
opt the adaptive cha
nge metho
d
to acq
u
ire the
inertia facto
r
and lea
r
nin
g
factor.
Referen
ces
[1]
J Laffert
y
,
A M
c
Callum, F Per
e
ira.
Conditional random
fi
elds: Probabilistic m
o
dels for segm
enting and
lab
e
li
ng se
que
nce data.
Proc.
18th Internati
o
nal C
onf On Machi
ne L
earn
i
n
g
. 282-2
89.
[2]
AR Kinjo, F
Rossell
o, G Valiente. Profile C
ond
it
ion
a
l Ra
n
dom F
i
elds for
Modeli
ng Prot
ein F
a
mil
i
es
w
i
t
h
Structural
Information.
BIOPHYSICS
. 2009; 5: 37-
44.
[3] B
Settles.
Biome
d
ic
al na
med
entity recogniti
on usin
g con
d
it
ion
a
l ran
d
o
m
fields an
d nov
el feature sets.
In Proceedi
ng
s of
the Joint W
o
rkshop on
Natura
l Lan
g
uag
e Processi
ng in Biome
d
i
c
ine an
d its
Applications (JNL
PBA-2004). 104–107.
[4]
M Bundsc
hus, M Dejor
i
, M Stetter, V
T
r
esp, HP Kr
ieg
e
l, E
x
tractio
n
of se
mantic bi
ome
d
i
cal re
latio
n
s
from text usin
g
conditi
on
al ran
dom fiel
ds, BMC Bioi
nformatic
s
9 (2008): 2
0
7
.
[5]
F
l
etcher R an
d
CM Reeves. F
unctio
n
Minim
i
za
tion b
y
C
onj
ugate Grad
ie
nts. Comp. 1964;
7:149-1
54.
[6]
M Al-Baa
li. De
scent pro
pert
y
and
gl
oba
l co
n
v
erge
nc
e
of the F
l
etcher-
R
ee
ves metho
d
w
i
t
h
in
e
x
act l
i
n
e
search.
IMA Journa
l of Nu
mer
i
cal An
alysis
. 1
985; 5: 12
1-12
4.
[7]
YH Dai
an
d
Y Yuan. C
o
n
v
erge
nce pr
op
erties of the
F
l
etcher-R
eev
es metho
d
.
IMA Jo
u
r
na
l
of
Nu
meric
a
l An
al
ysis
. 1996; 1
6
: 155-
164.
[8]
Charl
e
s A Sutton. Efficient T
r
aini
ng Meth
ods
F
o
r C
onditi
on
al Ra
ndom F
i
e
l
ds. PhD thesis,
Universit
y
of
Massach
usetts Amherst. 2008
.
[9]
Richar
d H B
y
r
d
, Peihua
ng L
u
,
Jorge Noced
a
l and Ci
yo
u Z
hu, A Limited Memor
y
Al
gor
ithm F
o
r Bound
Constra
i
ne
d Optimizati
on.
SIAM Journa
l on
Scientific C
o
mputin
g
. 199
5; 16: 1190-
12
08.
[10]
T
homas G D
i
etterich, Guo
hua H
ao, Ad
am As
henfe
l
ter. Gradient
T
r
ee Boosting
for
T
r
ainin
g
Con
d
itio
nal R
a
ndom F
i
e
l
ds.
Journ
a
l of Mach
ine L
earn
i
n
g
R
e
searc
h
. 200
8; 9: 2113-2
1
3
9
.
[11]
C Sutton, A McCall
um.
Piece
w
ise pseud
olik
elih
oo
d for efficient traini
ng o
f
conditio
nal ra
ndo
m fiel
ds
.
Procee
din
g
s of
the 24th inter
n
ation
a
l co
nf
ere
n
ce on Mac
h
i
n
e lear
nin
g
. 200
7; 863-8
70.
[12]
SVN Vish
w
a
na
than, NN Schr
aud
olp
h
, MW
Schmidt, KP Murph
y
.
Accel
e
rated trai
ni
ng
of conditio
n
a
l
rand
o
m
fields w
i
th stochastic gradie
n
t meth
ods
. Proceed
in
gs of the 23 rd
In
ternatio
na
l Confere
n
ce o
n
Machi
ne Le
arn
i
ng, Pittsburg
h
,
PA. 2006; 969
-976.
[13]
Chris Pal, Ch
arles Sutton, and An
dre
w
Mc
Call
um. Sparse for
w
ard
back
w
a
r
d usi
n
g minimum
diver
genc
e b
e
a
ms for fast traini
ng
of con
d
itio
nal r
a
n
d
o
m
F
i
elds.
In I
n
ternati
o
n
a
l C
onfere
n
ce
o
n
Acoustics, Spe
e
ch, and Si
gn
a
l
Processi
ng (ICASSP)
. 2006;
v-v.
[14]
J Kenn
ed
y a
n
d
RC Eber
hart. Particle S
w
a
r
m
Optimization.
Proc. on fee
d
b
a
ck mecha
n
is
m
IEEE Int
’
l.
Conf. on Ne
ura
l
Netw
orks, IEEE Service Ce
nter
. 1995; VI: 194
2-19
48.
[15] Y
Guang
yo
u.
A mo
difed
particl
e sw
arm o
p
ti
mi
z
e
r
alg
o
rith
m
. 8th
Internatio
nal
Confere
n
ce
o
n
Electron
ic Mea
s
ureme
n
t and I
n
strument
s. 20
07; ICEMI '
07:
2-67
5-2-6
79.
[16]
H anH, G iles
C, M anavo
g
lu
E, et al.
Automatic
Docu
me
nt Metadata Ex
traction Usi
ng
Supp
ort Vector
Machi
nes
. In: Procee
din
g
s o
f
Joint Confere
n
ce on D
i
git
a
l Libr
aries. 20
03
: 37 - 48.
[17] Laffert
y
J, M c
C
all
u
m A, Pere
ira F
.
C
o
nd
i
t
i
o
n
a
l
R
a
nd
om
Fie
l
d
s
: Prob
ab
i
listi
c Mo
d
e
l
s
fo
r Se
gm
en
ti
ng
and La
bel
in
g Sequ
enc
e Data
. In: Proceed
ings of the 18th Inte
rnation
a
l
Confere
n
ce on Machi
n
e
Lear
nin
g
. 200
1
:
282 - 289.
[18]
B
y
rd R
H
, Noc
eda
l J, Schna
bel RB. Re
pre
s
entatio
ns of Quasi-N
e
w
t
on
Matrices and
T
heir Use i
n
Limited Memor
y
Methods.
Mathematical Programm
ing
. 199
4; (2): 129- 15
6.
[19]
Darroch JN, Ratcl iff D. Gen
e
raliz
ed
Iterati
ve Scalin
g for
Log- li
near Mo
dels.
Anna
ls of Mathematica
l
Statistics
. 1972
; 43(5): 147
0 - 148
0.
[20]
Dell
a Pietra S,
Dell
a Pietra V
,
Laffe
rt
y
J. Inducin
g F
eatures
of Rand
om F
i
elds.
IEEE Transactions
on
Pattern Analys
i
s
and Mach
in
e Intelli
genc
e
. 19
97, 19(4): 3
80-
393.
[21]
Peng F
,
M cC
allum A. Accurate Informatio
n
Ex
tractio
n
from Researc
h
Papers Usi
ng
Conditio
n
a
l
Ran
dom F
i
el
ds
.
Informatio
n Processi
ng & Mana
ge
me
nt
. 20
06; 42(4): 9
63-
979.
[22]
Sha F
,
Pereira F
.
Shallow
Parsin
g w
i
th
Con
d
itio
nal R
a
ndo
m F
i
el
ds
. In: Proceed
in
g
s of Hum an
Lan
gu
age T
e
chno
log
y
NA
AC
L. 2003: 1
34-1
41.
[23]
Ruggieri S. Efficient C4.5.
IEEE Transactio
n
s
on Kn
ow
led
g
e
and
Data E
ngi
neer
ing
. 2
0
0
2
, 14(2): 4
38-
444.
[24]
Z
a
ve P. Classification of Research
Effort s in Requ
ireme
n
t’s Engin
eeri
ng.
ACM Computi
ng Surveys
.
199
7; 29(4): 31
5-32
1.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1213 – 1
220
1220
[25]
Yu L, Liu H. Efficient
feature selecti
on via a
nal
ysis
of relev
ance an
d redu
nda
nc
y
.
Jour
n
a
l of Machin
e
Lear
nin
g
Res
e
arch
. 200
4; 5(1
0
): 1205
−
12
24
.
[26] S
Das.
F
ilters, W
r
appers and
a
Boosti
ng-B
a
s
ed Hy
brid for
F
eature Se
lecti
o
n
. In: Proceedings of 18th
Confer
ence
on
Machin
e Le
arnin
g
. 200
1: 74-
81.
[27]
Pur
w
o
harj
o
n
o
, Abdilla
h, Muh
a
mmad, Pena
ngsa
ng, Ontoseno, Soepr
ija
n
t
o, Adi.
Optimal plac
ement
and sizi
ng of th
y
r
istor-c
ontro
lleds
eri
e
s-cap
a
c
it
or using gr
a
v
itation
a
l sear
ch algor
ithm.
Te
lkom
n
i
ka
.
201
2, 10(5): 89
1-90
4.
[28]
Sari, L
y
d
i
a. Effects of punctur
i
ng p
a
tterns o
n
punctur
ed c
o
n
v
oluti
ona
l co
de
s.
T
e
lkomnik
a
. 201
2; 10(
4):
752-
762.
Evaluation Warning : The document was created with Spire.PDF for Python.