TELKOM
NIKA
, Vol. 13, No. 4, Dece
mb
er 201
5, pp. 1456
~1
465
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i4.2300
1456
Re
cei
v
ed
Jul
y
8, 2015; Re
vised Septem
ber
23, 20
15;
Accept
ed O
c
tober 10, 20
1
5
Classification of Non-Functional Requirements Using
Semantic-FSKNN Based ISO/IEC 9126
Denni Aldi Ramadhani*
1
, Siti Rochimah
2
, Umi Laili
Yuhana
3
1
Informatics
Department, Fac
u
lty of Comput
er S
c
ien
c
e, Dia
n
Nu
swanto
r
o University
Imam Bonjol, Semara
ng 50
131, Indon
esi
a
2,3
Informatics
Dep
a
rtme
nt, Faculty of Informat
io
n Technolo
g
y, Institut Teknolo
g
i Sepuluh
Nop
e
mbe
r
, Surab
a
ya 601
1
1
, Indone
sia
*Co
rre
sp
ondi
ng autho
r, email: denni.al
d
i@d
s
n.di
nu
s.ac.id
1
; s
i
ti@
i
ts
-s
by.edu
2
;
yuhana
@if.its.ac.id
3
A
b
st
r
a
ct
Non-fu
nction
al
requir
e
ments
is one
of the i
m
p
o
rtant facto
r
s that pl
ay a role in th
e suc
c
ess of
softw
are deve
l
op
me
nt that is
often
overl
o
o
k
ed
by d
e
vel
o
pers, so
it cau
s
e a
d
verse
effects. In ord
e
r
to
obtai
n the no
n
-
function
al re
q
u
ire
m
e
n
ts, it requ
ires a
n
ide
n
tificatio
n
auto
m
ati
on syste
m
of non-functi
o
nal
requ
ire
m
e
n
ts. This resear
ch pro
poses
an a
u
to
mati
on syste
m
of
ide
n
tificatio
n
of non-fu
nction
a
l
requ
ire
m
e
n
ts from th
e req
u
ire
m
e
n
t sente
n
ce
-base
d
classifi
cation a
l
g
o
rith
ms of F
SKNN
w
i
th the additi
o
n
of
semantic f
a
cto
r
s such
as the
term
dev
elo
p
m
e
n
t by
hip
e
r
n
i
m
a
nd
meas
ure
m
e
n
t of se
ma
ntic re
late
d
ness
betw
een th
e te
rm a
nd ev
ery c
a
tegory
of qu
al
ity aspects
b
a
s
ed ISO / IEC 9126. In the test
, the dataset is
134
2 se
ntenc
e
s
from six
diff
erent d
a
tasets
.
T
he result
o
f
this researc
h
is that the S
e
mantic-F
SKN
N
meth
od c
an r
e
duce th
e va
lue
of ha
mmi
ng l
o
ss or error r
a
te
by 21.9
%
, an
d
also r
a
ise t
he
valu
e of acc
u
ra
cy
by 4
3
.7%,
and
als
o
the
pr
eci
s
ion
val
ue
a
m
ounte
d
to
73.
9
%
co
mpare
d
t
o
F
SKNN
met
hod
w
i
thout th
e
add
ition
of semantic factors in
it.
Ke
y
w
ords
: No
n-F
unctio
nal R
equ
ire
m
e
n
ts, Classificati
on, S
e
mantic-F
SKN
N, ISO/IEC 9126
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Knowin
g
the
non-fu
nctio
n
a
l
re
quirement
s clo
s
ely rela
ted
to softwa
r
e quality
a
s
pect
s
i
s
an essential
thing becau
se the quality aspe
ct
of non-fun
c
tion
al
requireme
nts is one of the
factors that
pl
ays a
role
in t
he
succe
s
s o
f
a syste
m
de
velopment. In
dustri
a
l
worl
d
,
in practi
ce, i
s
also
too fo
cu
sed
on
the f
unctio
nal
req
u
irem
ent
s fa
ctors
and
often forgets th
e facto
r
of n
on-
function
al re
q
u
irem
ents to
achi
eve the d
e
sig
n
and
de
velopment p
h
ase [1]. In fact, if the quali
t
y
asp
e
ct
of n
o
n
-fun
ctional
requireme
nts i
s
n
o
t ta
ke
n
i
n
to con
s
ide
r
a
t
ion an
d
not
kno
w
n, it
wo
uld
c
a
us
e
h
a
r
mfu
l
e
ffe
c
t
s
,
bo
th
fo
r
th
e
s
u
c
c
e
s
s
o
f
th
e
s
o
ftwa
r
e s
y
s
t
e
m
d
e
ve
lo
p
m
e
n
t
and
stakehol
ders.
One
ca
se i
s
the L
ond
o
n
Ambula
n
ce System (LAS), the L
AS system
failed in
perfo
rman
ce
due to un
ce
rtainty and in
consi
s
ten
c
y in
the pro
c
e
s
s o
f
non-fun
c
tion
al req
u
ire
m
en
ts
spe
c
ification [
2
]. Another
system failure is the he
alth
system of Ele
c
troni
c
Healt
h
Re
co
rd (E
HR)
that failed in
perfo
rman
ce
due to the
la
ck
of qua
lity
asp
e
ct
s of u
s
ability [3]. Other
effects
al
so
have an imp
a
c
t for sta
k
e
h
o
l
ders, as
hap
pene
d in
the developm
ent of
the
US Army Intelligen
ce
Sharing Syst
em that the
devel
opm
ent
cost up to $ 2.7 billion
wasted because
the system
is
con
s
id
ere
d
u
s
ele
s
s beca
u
se of probl
ems in c
apa
city factor a
nd the qualit
y factor of the
perfo
rman
ce
and u
s
ability [4].
Some cases
above a
r
e
pictures sho
w
in
g that
no
n-fu
nction
al requi
reme
nt is
an i
m
porta
nt
factor
whi
c
h
must be kno
w
n and
ident
ified
fi
rst
bef
ore th
e
software
develo
p
m
ent ente
r
s
the
advan
ced p
h
a
se. Howeve
r, identifying non-fu
nctio
n
a
l
requi
rem
ent
is not an e
a
sy thing becau
se
of several factors
su
ch a
s
the lack of sta
ndards
in i
d
e
n
tifying the non-fun
c
tion
al requi
rem
ent, the
non-fu
nctio
n
a
l
req
u
ire
m
ent
is often
en
countered i
n
complete
and
it is often
hid
den o
r
mixe
d
in
the fun
c
tiona
l req
u
ire
m
en
t sente
n
ces
[5] [6], t
he requireme
nt senten
ce
writt
en in
natu
r
a
l
langu
age u
s
u
a
lly have am
biguity whi
c
h
woul
d ma
ke it
difficult to identify the aspe
cts of qu
ality of
non-fu
nctio
n
a
l
requi
rem
ent
contai
ned i
n
it [7]. Ther
ef
ore,
we n
eed
ways to
be a
b
le to ide
n
tify
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No
. 4, Decem
b
e
r
2015 : 145
6 – 1465
1457
asp
e
ct
s ofthe
quality ofn
o
n
-
functio
nal
re
quire
ment,
o
n
e
of
whi
c
hi
s b
y
having th
e
cla
ssifi
cation
of
the sente
n
ce
s of req
u
irem
ent wri
tten in
the requi
rem
ent document
.
This
re
sea
r
ch aim
s
to d
e
s
ign
a
syste
m
that
can a
u
tomatically
cla
ssifyn o
n
-f
unctio
nal
requi
rem
ent f
r
om va
riou
s
requi
rem
ent
document
s.
This
syste
m
will have
three mai
n
p
h
a
s
e
s
namely a
u
to
mation of t
r
aining
data
labeling,
me
asu
r
em
ent o
f
sema
ntic
relatedn
ess,
and
cla
ssif
i
cat
i
on
pro
c
e
s
s
co
mpri
sing
t
h
e
st
ep
s
of
training
an
d classificatio
n
. The pro
c
e
ss
of
automation
o
f
training
dat
a labeli
ng i
s
done i
n
o
r
de
r to
save m
o
re time th
an
if the labeli
n
g is
done ma
nual
ly, this process is do
ne
by TF-IDF
we
ighting
with
a sea
r
ch for the deg
ree
of
simila
rity usi
ng the co
sin
e
mea
s
urea
s done by
Su
harso a
nd Rochim
ah [8]. The pro
c
e
s
s of
sema
ntic rela
tedne
ss mea
s
ureme
n
t
is done by
usi
n
g the meth
o
d
of HSO [9]
[10] to get the
sema
ntic rel
a
tionshi
p between ea
ch cl
a
ss a
nd ea
ch
term that will be pro
c
esse
d. While for the
cla
ssif
i
cat
i
on
pro
c
e
ss,
it
u
s
e
s
FS
K
N
N
met
hod
s
that
have bee
n i
n
trodu
ce
d by
Jian
g et al [11]
that there is a new vari
able additio
n
in the tr
aining pha
se so
that the value of sema
n
t
ic
related
n
e
ss o
b
tained in the
previou
s
pro
c
e
ss
can
b
e
take
n into co
n
s
ide
r
ation d
u
ring the proce
s
s
of training.
Cla
s
ses i
s
u
s
ed fo
r the pro
c
e
ss
of
classificatio
n
is ba
sed o
n
the intern
ational
stand
ard ISO
/IEC9126.
There a
r
e m
a
ny method
s d
e
velope
d to d
o
cla
s
s
i
fic
a
tion process
both for documents
in
gene
ral a
n
d
requi
rem
ent
document
s to identify non-fu
nctio
nal re
qui
rem
ent of software,
con
s
i
s
ting of
the formatio
n
of NFR
Lo
cator sy
stem who
s
e cla
s
sification uses KNN
a
nd
ma
ke
s
use
of the function
of dist
ance Leven
shtein [12],
Naïve Bayes with some
dev
elopme
n
ts [6]
[13]
[14], using S
V
M method
with the re
ne
wal in feat
u
r
e
sele
ction p
h
ase [15], the
usa
ge of TF
-IDF
weig
hting an
d mea
s
u
r
eme
n
t of simila
rity degree of
co
sine m
e
a
s
u
r
e
[8]. Basically
, the pro
c
e
ss
of
cla
ssifi
cation
con
s
i
s
ts
of si
ngle-l
abel
an
d multi-l
abel,
so
me m
e
th
od d
e
velopm
ents
mention
e
d
above
are
sti
ll limited in
si
ngle-l
abel,
where
a
s in
fa
ct, there i
s
a
possibility th
at one
option
of
requi
rem
ent
sente
n
ce in
req
u
ire
m
ent
do
cume
nt contai
ns more
than one
aspe
ct
of non-
function
al re
quire
ment (
m
u
lti-label
) [7] [16]. The method dev
elopme
n
t ab
ove, has al
so not
con
s
id
ere
d
a
facto
r
of sema
ntic rel
a
tionshi
p
s
b
e
twee
n ea
ch
cate
gory
of non
-fun
ctio
nal
requi
rem
ent and term tha
t
are pro
c
e
s
sed duri
ng th
e
training, be
cause in fact sema
ntic fact
or
can imp
r
ove t
he cla
s
sificati
on re
sult to be better [17] [18].
One research trying to perform m
u
lti-l
abel cl
assification of documents
wa
s
done by
Jun
g
-Yi
Jia
n
g
et al
that
p
r
opo
se
d a
m
e
thod
ca
ll
ed
Fuzzy Simila
rity base
d
K-Nea
r
e
s
t
Neig
hbor
(FSKNN) [11
]. FSKNN method is p
r
o
v
ed to be better than other multi-l
a
b
e
l cla
ssifi
cation
method
s, but
the FSKNN method ha
s not taken
into accou
n
t the sema
ntic facto
r
s in
it.
Therefore, thi
s
re
se
arch wi
lluse the m
e
thod FSKN
N
by adding
se
mantic fa
ctor
in the form of
ter
menri
c
hm
ent
in the trainin
g
of databa
sed on
co
mbi
nation correl
ation betwee
n
hypernyms and
synonym
s
ba
sed
WordNet and sema
ntic relate
dne
ss measu
r
em
e
n
t between t
he term and
the
categ
o
ry of non-fun
c
tion
al requi
rem
ent whi
c
h is a
r
e n
e
wal fa
ctor gi
ven in this re
sea
r
ch.
2. Rese
arch
Metho
d
Overall, the d
e
sig
n
of the sy
stem a
r
e
made ba
se
d on the re
se
a
r
ch by Jia
ng
et al [11]
are
sh
own in
Figure 1
with
dark-colored
part
is
a
con
t
ribution i
n
thi
s
resea
r
ch. Base
d o
n
Fig
u
r
e
1 above, this rese
arch m
e
thod con
s
ist
s
of four
mai
n
pha
se
s, namely: automation of traini
ng
data lab
e
ling,
sem
antic
rel
a
tedne
ss me
asu
r
em
ent, Semantic-FSK
NN t
r
aini
ng p
hase con
s
isti
ng
of trainin
g
pat
tern g
r
ou
ping
and
cal
c
ulati
on of
p
r
io
r p
r
obability an
d l
i
kelih
ood
s val
ue, and
the la
st
is Semanti
c
-F
SKNN cl
assifi
cation p
h
a
s
e.
2.1. Automa
tion Labeling
Training Data
On the pha
se of the automation lab
e
ling tr
ainin
g
data co
nsi
s
t
s
of four ph
ase
s
:
prep
ro
ce
ssin
g, term en
richment in the
trai
ning data based
on co
mbinati
on co
rrelation between
hypernym
an
d syn
onym
s
,
the weightin
g
of tf-idf,
an
d s
i
mila
r
i
ty va
lu
e
me
as
u
r
eme
n
t. All ph
as
es
of automatio
n lab
e
ling t
r
ai
ning
data i
s
based
on
re
search th
at ha
s b
een
don
e
by Suha
rso a
n
d
Ro
chima
h
[8]
.
Except in
seco
nd p
h
a
s
e
is the fi
rst
contributio
n of
this resea
r
ch that is to
e
n
rich
relevant term
for training data usin
g a
comb
in
ation
pattern hypernym and
synonyms a
s
in
Figure 2.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Cla
ssifi
cation
of Non-F
u
n
c
tional Requi
re
m
ents Usin
g Sem
antic- …
(Denni Aldi Ram
adhani
)
1458
2.2. Semantic Relatedne
ss Meas
ure
m
ent
This ph
ase
is se
cond
contrib
u
tion
in
this research. Mea
s
urem
ent of sema
ntic
related
n
e
ss i
n
this research
will use the metho
d
o
f
Hirst & S-O
nge (HSO
). In the metho
d
of
HSO, se
man
t
ic relate
dne
ss ca
n be m
e
asu
r
ed
by
graph of vocab
u
lary contain
ed in Word
Net
whi
c
h
con
s
i
s
ts of n
ode
s th
at rep
r
e
s
ent
words an
d rel
a
tions
amo
n
g
node
s th
at d
e
scrib
e
different
relation
shi
p
s.
Base
d o
n
the g
r
ap
h co
nce
p
ts, the
method
of HSO mea
s
u
r
e
s
the
se
man
t
ic
related
n
e
s
s
usin
g the
pat
h di
stan
ce b
e
twee
n bot
h
word n
ode
s
(path di
stan
ce
), a n
u
mbe
r
of
cha
nge
s of di
rectio
n of the path co
nne
cti
ng bot
h word
node
s an
d ba
sed o
n
the all
o
wa
ble path.
Figure 1. System De
sign
Figure 2. Pattern of Co
mbi
ned Develop
m
ent of Hipe
rnim and Syno
nyms
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 13, No
. 4, Decem
b
e
r
2015 : 145
6 – 1465
1459
Basic ide
a
of the
HSO
me
thod i
s
to
det
ermin
e
the
semantic relat
edne
ss b
e
tween t
w
o
wor
d
s t
h
at
is
comp
are
d
u
s
ing
coh
e
si
o
n
relat
i
on
s to
calculate th
e allowable p
a
th betwe
en
two
words. T
he
HSO method
h
a
s three diffe
rent types
of coh
e
si
on rela
tions
whi
c
h di
rectly conn
ects
the sema
ntic
related
n
e
ss b
e
tween two words
[9] [10].
2.3. Extra Str
ong Rela
tion
Extra strong relation is a re
lation betwe
e
n
two
words t
h
at have the hi
ghe
st weig
ht of all
kind
s of othe
r relation
shi
p
s and ge
nerates hig
h
corr
el
ation. An Example of extra
stron
g
rel
a
tion
is su
ch
com
pari
s
on of th
e same
wo
rds that are
“man” a
n
d “man” [10]. Value of sem
a
ntic
related
n
e
ss f
o
r extra st
ron
g
relation i
s
o
b
tained fro
m
equatio
n
:
2*C
(
1
)
whe
r
e C i
s
a fixed con
s
tant
by 8
.
2.4. Strong Relatio
n
Strong
Rel
a
tion
will o
c
cur i
n
two
word
s t
hat have
the
same
pa
re
nt
word o
r
deriv
ed fro
m
the same pa
rent. Relation
use
d
in the concept of
parent word is b
a
se
d on the relation of IS-A.
Two
wo
rd
s a
r
e said to h
a
v
e stron
g
rel
a
tion by
the
followin
g
con
d
itions
(a
) When two words
sha
r
e th
e sa
me pa
rent
concepts.
(b)
Whe
n
the
r
e
are
asso
ciati
on relation
s i
n
the form
o
f
a
hori
z
ontal
lin
k
(a
ntonyms, similar
to, se
e also, attr
ibute
)
b
e
twee
n pa
rent wo
rd of both wo
rds.
F
o
r
example, the
word m
an a
nd wom
an h
a
ve a corr
ela
t
ion of strong
relation be
cause both ha
ve
hori
z
ontal li
nks in the fo
rm
of an anto
n
ym. (c) Wh
en t
h
ere i
s
a
n
y link b
e
twe
en t
he pa
rent
word of
each wo
rd, if one wo
rd i
s
a compo
und
word o
r
phr
ase
that include
s other wo
rd
s, For example
,
the wo
rd colo
r and
wate
r-color. To me
a
s
ure the se
m
antic related
n
e
ss in Stron
g
Relation i
s
the
same a
s
in e
x
tra stron
g
rel
a
tion, it uses
equatio
n 5.
2.5. Medium Strong Relation
In a medium stron
g
rel
a
tio
n
, semanti
c
rela
tedn
ess m
easure
m
ent i
s
don
e by co
nsid
erin
g
paths
allowe
d (allo
wa
ble
path) a
nd nu
mber
of cha
n
ge of directio
n. Path Detail
allowe
d an
d not
allowed can b
e
see
n
in Fig
u
re 3.
Figure 3. (a)
Allowabl
e pat
h (b) Pattern of unallo
wabl
e path [10]
HSO m
e
thod
provid
es two rul
e
s to e
n
su
re
th
at a
path
w
ay i
s
approp
riate
with th
e
relation
shi
p
b
e
twee
n a so
u
r
ce a
nd a target word, as f
o
llows:
Rule 1: there
is no link th
at precede
s
upw
ard lin
k.
Once a wo
rd
is narro
wed
down by
usin
g the li
nk downward o
r
hori
z
o
n
tal lin
ks, it i
s
n
o
t al
lowe
d to g
e
n
e
rali
ze th
e word
agai
n
u
s
i
ng
upward lin
k.
Rule
2: at
mo
st, only o
ne
chang
e of
dire
ction
i
s
allo
wed.
An act
to
cha
nge
the di
rectio
n
is a
big
ste
p
i
n
the
dete
r
mi
nation
of the
sema
ntic
rel
a
tedne
ss bet
w
een t
w
o
wo
rd
s, the
r
efore, the
cha
nge in di
rectio
n sh
oul
d be limited. But there ar
e two exce
ptions to this
rule, whi
c
h it is
permitted to u
s
e a ho
rizont
al link to make the tran
sition from the top downward
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Cla
ssifi
cation
of Non-F
u
n
c
tional Requi
re
m
ents Usin
g Sem
antic- …
(Denni Aldi Ram
adhani
)
1460
In the m
ediu
m
strong
rela
tion, in o
r
d
e
r to
me
asure
the sema
ntic rel
a
tedne
ss
betwe
en
two wo
rd
s of HSO metho
d
, it uses the fo
llowing
equati
on:
∗
/2
∗
(2)
Whe
r
e
C is
a
fixed con
s
ta
nt by 8,
k are
fixed con
s
ta
nts by 1. Whil
e
Pl
is the p
a
t
h length
and
Nd
i
s
the
numbe
r
of
ch
ange
di
re
ction. the val
ue
o
f
(2
*
C) is u
s
ed fo
r n
o
rm
ali
z
e th
e
sem
a
n
t
ic
related
n
e
ss v
a
lue to b
e
in
a ran
ge of 0
to 1 t
hat will
be re
qui
red
by the metho
d
of Semanti
c
-
FSKNN.
2.6 Semantic
-FSKNN T
r
ai
ning Phase
Semantic-FS
K
NN traini
ng
phase
con
s
i
s
ts of two ph
ase
s
: gro
upin
g
pattern
s of training
data and p
r
io
r pro
bability a
nd likeli
hood
s calculation [1
1].
2.7. Training Patter
n
Gro
uping
Grou
ping the
training do
cuments
,
,…,
into
p
cluste
rs b
a
se
d on
fuzzy
s
i
milar
i
t
y
m
easure.
It is given
,
and
,
that is the distributio
n of
term
at the category
,
that is define
d
as follo
ws:
,
∑
∑
(3)
,
∑
∑
(4)
For 1
≤
i
≤
nu
mber of term
and 1
≤
j
≤
nu
mber of cl
ass or cate
gory, whe
r
e :
1,
0
0,
0
(5)
So that it will surely obtain
0
≤
,
and
,
≤
1. The next ste
p
s i
s
to calcul
ate
the deg
ree
of membe
r
sh
ip
on categ
o
ry
, in the pro
c
e
ss
of calcul
ating th
e deg
ree of
membership
on catego
ry
it is given a
new fo
rmul
a
in ord
e
r th
at value obtai
n
ed from th
e
measurement
pro
c
e
s
s of
sema
ntic
rel
a
tedne
ss
can
be ta
ken
into
acco
unt. Th
e ad
dition
of a
formula
conta
i
ned in the pa
rt in bold (
Rel(ti,cj
)
)
, as
follows
:
,
,
,
,
,
,
,
,
,
,
(6)
For 1
≤
i
≤
n
u
mbe
r
of term and 1
≤
j
≤
numbe
r of
cl
ass or
cate
g
o
ry. Whe
r
e
Rel(ti,c
j)
is
the value of sema
ntic rel
a
tedne
ss bet
ween term
t
i
an
d the catego
ry
c
j
obtaine
d at the phase
of
sema
ntic
rela
tedne
ss mea
s
ureme
n
t. Every value
of
Rel(ti,c
j)
will
be divided by
the highest value
of the overall value of
Rel(tu,c
v)
.
The next phase is to determine
the fuzzy
similarity of each do
cume
nt d,
d=
〈
,
,…,
〉
on categ
o
r
y
as
follows
:
,
∑
,
⨂
∑
,
⨁
(
7
)
Whe
r
e
⨂
and
⨁
is
fuzz
y
t-norm
and
t-c
o
norm
which is su
bse
que
ntly defined a
s
follows :
⨂
(8)
⨁
(
9
)
And
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No
. 4, Decem
b
e
r
2015 : 145
6 – 1465
1461
(10
)
it is the degree of membe
r
shi
p
of the te
rm of the do
cume
nt. The final phase is to
define the de
gree of mem
bership of a d
o
cum
ent
d
to the c
a
tegory
as
follows
:
,
,
(
1
1
)
For 1
≤
j
≤
n
u
m
ber
of cl
ass or
categ
o
ry.
For e
a
ch trai
ning d
o
cume
nt
, 1
≤
i
≤
n
u
m
ber of
document
,
wi
ll have the
calcul
ation
, 1
≤
j
≤
nu
mber
o
f
c
l
a
s
s
or c
a
te
go
r
y
, us
in
g
the
equatio
n 15. To define
p
clu
s
t
e
r,
,
,…
,
,
is
as
follow:
|
,
1
}
(
1
2
)
For 1
≤
v
≤
nu
mb
er
o
f
c
l
ass
or
ca
te
g
o
r
y
,
whe
r
e
α
i
s
a thre
sh
old d
e
fined by th
e
use
r
fo
r
use i
n
the training
pro
c
e
s
s.For
every
, 1
≤
i
≤
num
b
e
r of d
o
cume
nt
,
is define
d
as
sea
r
ch set
for every
⊆
if and only if
∈
, 1
≤
v
≤
num
ber o
f
class or cat
egory.Th
e
ne
xt
process wil
l
be de
scri
bed
in p
s
eu
do-code
sho
w
n
in
Figure 4,
with a n
o
te in t
he be
ginni
ng
∅
, 1
≤
v
≤
numbe
r of cla
ss o
r
catego
ry
,
and
∅
, 1
≤
u
≤
numb
e
r of
document
.
Figure 4. Pse
udo-co
de g
r
o
uping p
r
o
c
e
s
s and t
he p
r
o
c
e
ss of defini
ng the se
arch
set G
i
[11]
Output that is in the form
of
sear
ch set
,
,…,
will then be use
d
to determine the
nearest n
e
igh
bor d
a
ta that can h
e
lp p
e
rf
orm
ca
lculati
on of pri
o
r p
r
obability valu
e and the val
ue
of likeliho
od in the next phase.
2.8. Calculati
on of Prior Probabilit
y
an
d Likelihoods Value
It is given
that is pri
o
r probability whose va
lue must
be known
before continui
ng
into every obse
r
vation, while
|
is a
cla
ss of likel
ihood an
d a
condition
al prob
ability
that
has be
en
linked with
observation
E. This pr
ob
ability calcul
a
t
ion is done
on the trainin
g
pattern
s obtai
ned previou
s
l
y
, as follows :
1
∑
(
1
3
)
0
1
1
(
1
4
)
Whe
r
e
s is a
smoothi
ng
co
nstant val
ue,
whi
c
h i
s
usua
lly a small
po
sitive real
wo
rth. The
next phase i
s
cal
c
ulatin
g likeliho
od cla
ss of
|
. E
can be
0,1,
... , or
k
. Fo
r every training
document
, 1
≤
i
≤
nu
mbe
r
of do
cume
n
t
,
where
=
〈
,
,…,
〉
is
k-ne
arest n
e
ighb
or
obtaine
d from
sea
r
c
h
set
dan
〈
,
,…,
〉
, which i
s
a label qu
antity vector define
d
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Cla
ssifi
cation
of Non-F
u
n
c
tional Requi
re
m
ents Usin
g Sem
antic- …
(Denni Aldi Ram
adhani
)
1462
∑
(
1
5
)
For 1
≤
j
≤
nu
mber of cl
ass or cate
gory. The next step
is defined :
,
∑
(
1
6
)
,
∑
(
1
7
)
whe
r
e
1
and
1,
0,
(
1
8
)
Then defin
es
likeliho
o
d
s
, as follows :
|
1
,
∑
,
(19
)
|
0
,
∑
,
(20
)
For eve
r
y
e
=
0, 1, ...
,
k
and
j
=
1, 2, ...
,
number of c
l
ass
or
c
a
tegory, bec
ause the s
i
z
e
of
, 1
≤
i
≤
nu
mber
of do
cu
mentisal
way
s
small
e
r th
an
the num
be
r
of initial traini
ng patte
rn
s,
comp
uting likelihoo
ds
can
be don
e effici
ently.
2.9. Semantic-FSKNN Cl
assifica
tion
Phase
Semantic-FS
K
NN metho
d
testing
p
r
o
c
ess i
s
d
one
by usi
ng th
e
estimated
ma
ximum a
poste
rio
r
i (M
AP). For example
,
,…,
is a set
k-n
e
a
r
e
s
t neighb
ors
for tes
t
ing
document
, a
nd
〈
,
,…
,
〉
is the vector sum label
for
(equ
atio
n 19), to determine
whi
c
h categ
o
ry that ha
s a rel
a
tion
ship
with
is by cal
c
ulatin
g the
label vect
or
〈
,
,…,
〉
from the do
cume
nt
by using e
s
timati
on
m
a
xim
u
m a posteri
ori
(MAP) as
follows
:
1,
1
0
0,
0
1
0,1
,
(21
)
For 1
≤
j
≤
n
u
mbe
r
of
cl
a
ss
or
cat
e
g
o
r
y
,
where
is a
rand
om vari
able to d
e
termine
entry into a
category
or n
o
t (
1
foryes, an
d
0
for n
o
),
E
is a variabl
e fo
r the n
u
mbe
r
of
document
s with
associ
ated with th
e cate
gory
, and
R
[0,1] indicates
0 or 1
cho
s
e
n
rand
omly. By
usin
g
Bayes
Rule
it is obta
i
ned:
(22
)
for b = 0, 1. Therefo
r
eth
e
e
quation 2
1
will
chan
ge into :
1,
1
1
0
0
0,
0
0
1
1
0,
1
,
(23
)
For 1
≤
j
≤
p.
To cal
c
ulate
i
t
m
u
s
t
g
e
t
, and cal
c
ulat
e
(equatio
n 13an
d 14
)
and al
so
|
(Eq
uation 1
9
an
d 20
). The
ca
lculatio
ns th
a
t
doesnot d
e
pend
on
, it can be
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No
. 4, Decem
b
e
r
2015 : 145
6 – 1465
1463
done o
n
the trainin
g
proce
ss a
nd the re
st is don
e wh
en the cla
s
sification p
r
o
c
e
s
s. The follo
wi
ng
is pha
se
s du
ring the cla
s
sification p
r
o
c
e
s
s
:
1. Cal
c
ulate
, 1
≤
j
≤
p
by e
q
u
a
tion 11
u
s
in
g the info
rma
t
ion or th
e re
sults of calcul
ations
perfo
rmed by
Equation 6, 7 and 10.
2. Che
c
k
if
for 1
≤
j
≤
numbe
r of class or categ
o
ry
,
then it will get
search set
for
that is
∪
∪…
∪
.
3.
Find the set
from
k-nea
re
st neighb
ors
on
from
, and get the amount of label vecto
r
.
4. Cal
c
ulate
, 1
≤
j
≤
p
using equatio
n 23 by using the inform
ation that
has been cal
c
ulate
d
in
the training p
r
oce
s
s with Equation 1
3
, 14, 19and 2
0
.
5. if
that is obtained is 1, then
belon
gs cate
gory
, otherwi
se
doe
sn’tbel
ong to
.
3. Results a
nd Analy
s
is
Before an ex
planatio
n of the test re
sult
s and evalu
a
t
ion, will be explained a
b
out th
e
dataset use
d
in the testing
and expe
rime
nts setting in
this re
sea
r
ch.
3.1. Data
se
t
Testing
pha
se in this re
se
a
r
ch u
s
ing
a datasetfrom 134
2 re
quire
ment se
ntences
obtaine
d from
6 dataset
s that
Prom
ise
[19]
, Itrus
t, CCHIT, World Vis
t
a
US
Veterans Health Care
System
Do
cum
entation, Online Proje
c
t Mark
ing
System
SRS, Mars Expre
ss A
s
pe
ra-3
Processing and Archivi
ng
Facility S
R
S
. Whe
r
e the 1
3
42 se
nten
ce
s will be label
e
d
automati
c
all
y
with the step
s de
scribe
d in pr
eviou
s
sub-se
ction. F
r
om 13
42 se
ntences
woul
d be taken 6
0
%
(805
senten
ces) to be
u
s
e
d
as traini
ng
data an
d 40
% (537
sente
n
ce
s) u
s
ed f
o
r te
st data t
hat
whi
c
h will be
evaluated
with accuration, pre
c
isi
on, re
call, and ham
ming loss.
3.2 Experimental Setting
s
The exp
e
rim
ental p
h
a
s
ei
nthis
re
sea
r
ch is
do
ne
by a
scena
rio t
hat is by h
a
v
ing a
comp
ari
s
o
n
betwe
en the
origin
al FSK
N
N metho
d
and FSK
N
N
method th
at
have be
en
a
dded
with sem
anti
c
factors in
side
so-call
e
d Se
mantic-FSKNN met
hod. Every method in the
experim
ental scena
rio will be
te
sted usi
ng
u
s
in
g
the
para
m
eter a
m
ount
of ne
a
r
est
nei
ghbo
rs
(
k
)
from 1
0
to
5
5
an
d ea
ch
k
is te
sted
wit
h
a th
re
shol
d
of 0.1 to
0.9
.
For te
sting
at any thresh
old
value, it will be evaluat
ed by
4 metrics of ham
mingl
oss, accuracy, prec
i
s
ion, andrecall
evaluation.
3.3 Experimental Result
Based
on
the
analy
s
is of t
he te
st result
, it is
kno
w
n
that the b
e
st
total of
k
is 30
. T
h
e
Evaluation of the test result is sho
w
n
in T
able 1, whi
c
h sh
ows the value of Hamming lo
ss,
accuracy, pre
c
isi
on, and re
call.
Table 1. Tab
u
lation of ha
mming lo
ss,
accura
cy, pre
c
isi
on, re
call
based on FS
KNN meth
od
and
Semantic-FS
K
NN meth
od
with the amo
unt of k = 30 i
n
the scena
ri
o 2
Thres
hold
FSKNN Semantic-FSKNN
K=30 K=30
Hamm. Acc.
Pr
ec.
Recall
Hamm.
Acc.
Pr
ec.
Recall
0.1
32.3%
20.7%
44.7%
27.8%
31.9%
27.2%
45.6%
40.1%
0.2
33.6%
30.2%
45.8%
48.8%
30.4%
25.2%
53.7%
33.7%
0.3
34.3%
28.7%
44.6%
46%
30.5%
24.5%
50.7%
33.6%
0.4
29.3%
33.4%
51.6%
49.6%
25.4%
35.5%
64.4%
43.9%
0.5
26.3%
38%
57.2%
55.5%
24.7%
38.6%
65.4%
49.5%
0.6
23%
42.8%
63.6%
58.7%
25.3%
39%
63.7%
51%
0.7
23.6%
41.9%
63.6%
56.1%
22.3%
42.4%
73.9%
50.6%
0.8
23.5%
41.4%
67.7%
53.6%
21.9%
43.7%
68.7%
54.7%
0.9
22.9%
42.2%
68.1%
55.9%
22.5%
43.2%
68.8%
54.4%
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Cla
ssifi
cation
of Non-F
u
n
c
tional Requi
re
m
ents Usin
g Sem
antic- …
(Denni Aldi Ram
adhani
)
1464
Based o
n
th
e Comp
ari
s
o
n
of evaluation
between t
he FSKNN method an
d
Semantic-FS
K
NN
method in T
able 1, it can be see
n
that the
method of Seman
t
ic-FSKNN re
sults the lo
west
Hammi
ng lo
ss at 21.9% a
nd the hig
h
e
s
t accu
ra
cy
a
t
43.7%, whi
c
h b
o
th we
re
obtained
at the
threshold of
0.8, wherea
s the highe
st pre
c
isi
on val
ue is al
so ge
nerate
d
by Semantic-FSK
NN
method amo
unted to 73.9
%
obtained in the thresho
l
d
0.7 but for recall, it is
kno
w
n that the
FSKNN meth
od i
s
highe
r t
han S
e
manti
c
-FSK
NN me
thod
with
a v
a
lue
of 5
8
.7
% obtain
ed i
n
the
threshold 0.6.
The d
e
cli
ne i
n
value
of h
a
mming
lo
ss
of Sema
ntic-FSKNN met
hod i
s
1.6%
and th
e
increa
se in value of
accu
racy of
Sema
ntic-FSK
NN i
s
2.3%, in the result com
pari
s
on at th
e
threshold
0.8
and
al
so
pre
c
isi
on val
ue
of Sem
anti
c
-FSKNN meth
od by
10.3%
in compa
r
i
s
o
n
of
the re
sults
at the thre
shol
d of
0.7, it indicates the p
e
rf
orma
nc
e of Semantic-FSK
NN at the th
ree
metrics a
r
e b
e
tter than FS
KNN, but for
there
call pe
rforman
c
e of
FSKNN meth
odis b
e
tter th
an
the
meth
od o
f
Semantic-F
SKNN with a
n
in
crease of
7.7% in th
e
compa
r
ison
re
sult in
a
thre
shold
0.6. Declin
e in performan
ce of re
call on t
he Sema
ntic-FSK
NN
method o
c
cu
rs be
ca
use the
pro
c
e
ss
of forming the t
r
ai
ning patte
rn
on Semanti
c
-FSKNN m
e
th
od have a
ddit
i
onal p
r
o
c
e
s
s of
sea
r
ch for
relation
ship
o
f
sema
ntic relatedn
ess b
e
twee
n ea
ch
categ
o
ry of
non
-functio
n
al
requi
rem
ent
and a li
st of relevant term
so that t
he d
a
ta formed fo
r the traini
ng
pattern i
s
filtered
mo
r
e
th
or
o
ugh
ly.
4. Conclusio
n
Based o
n
the test result in this rese
arch, it can be con
c
lu
de
d that the addition of
sema
ntic factors on the
FSKNN met
hod will im
p
r
ove the perf
o
rma
n
ce of hammin
g
lo
ss b
y
21.9%, 43.7% for the accura
cy and 73
.9% for t
he preci
s
io
n. Re
call declin
e in perfo
rman
ce
on
the method
o
f
Semantic-F
SKNN o
c
curs becau
se th
e
pro
c
e
ss
of trainin
g
patte
rn formatio
n o
n
Semantic-FS
K
NN metho
d
ha
s ad
ditional
pro
c
e
s
s of
search
for
se
man
t
ic related
n
e
ss
relation
shi
p
betwe
en e
a
ch cate
go
ry o
f
the non
-f
un
ctional
re
quirement a
nd a
list of relev
ant
terms
re
sultin
g that the data formed fo
r the tr
ainin
g
pa
ttern are m
o
re stri
ctly filtered.
Referen
ces
[1]
S Ullah, M Iqbal and AM Khan.
A Survey on Issues in Non-
F
unction
al Re
q
u
ire
m
e
n
ts Elicit
ation
. 20
11
Internatio
na
l
C
onfere
n
ce on Comp
uter
Net
w
o
r
ks
a
nd Info
rmation T
e
chn
o
lo
g
y
(ICCNIT
)
. Pesha
w
a
r
,
Pakistan. 20
11
; 333-34
0.
[2]
A F
i
nkelstei
n and J Do
w
e
l.
A Come
dy o
f
Errors:
the Lon
do
n Amb
u
l
ance Serv
ice case study
.
Procee
din
g
s of
the 8th Int
e
rna
t
iona
l W
o
rksho
p
on S
o
ft
w
a
re
Specific
ation
a
nd D
e
sig
n
. Sc
hloss V
e
le
n
.
199
6; 2-4.
[3]
J Bertman and
N Skolnik. EH
Rs Get a Failin
g Grade on Us
abil
i
t
y
.
Internal Medicine News.
2010;43
:
50.
[4]
C Hosk
ins
on, "
P
o
litic
o
," N
e
w
s
, 29
J
uni
20
11.
[Onlin
e]. Avai
lab
l
e:
http://
w
w
w
.
p
o
lit
ico.com/ne
w
s
/
s
tories/06
11/5
805
1.html. [Accessed 2
0
Oktober 2
0
1
4
].
[5]
M Rahimi, M Mirakhor
li a
n
d
J Clela
nd-H
u
ang.
Auto
mate
d Extraction a
nd Visu
ali
z
at
io
n of Quality
Conc
erns fro
m
Re
quir
e
me
nts Specificat
ions
. 20
14 IEEE 22nd I
n
ternati
o
n
a
l
Req
u
irem
ents
Engi
neer
in
g C
onfere
n
ce (RE)
. Karlskrona, S
w
e
d
en. 20
14; 253 -2
62.
[6]
A Casam
a
yor,
D God
o
y
an
d M Cam
po.
Identific
ati
on
o
f
non-fu
nction
al re
quir
e
me
nts in te
xtu
a
l
specific
ations :
A semi-s
uper
vised l
earn
i
ng
appr
oach.
Info
rmati
on a
nd S
o
ftw
are T
e
chnolo
g
y.
201
0;
52: 436-
44
5.
[7]
A Rash
w
a
n,
O Ormandjieva and R Witte.
Ontology-
B
ased
Class
ificatio
n of No
n-F
unctio
n
a
l
Req
u
ire
m
ents i
n
Softw
are Sp
ecificati
ons:
A
new
Cor
pus
a
nd SVM-B
ased
Class
ifier
. 2013 IEEE 37t
h
Annu
al Com
p
u
t
er Soft
w
a
r
e
an
d Appl
icatio
ns Confer
ence.
K
y
ot
o. 201
3; 38
1-38
6.
[8]
W
Suharso
a
n
d
S R
o
chim
ah
. Sistem Detek
s
i da
n Kl
asifik
asi Otomatis K
ebutu
h
a
n
N
o
n
F
ungsi
o
n
a
l
Berbas
is ISO/IEC 912
6. Master T
hesis. Suraba
ya: Postgr
ad
uate IT
S; 2013
.
[9]
M Cho
u
d
hari.
Exte
ndi
ng T
h
e
Hirst An
d St-
O
nge
M
easur
e
Of Semantic
Relat
edn
ess F
o
r T
he Un
ified
Medic
a
l La
ng
u
age S
y
stem. Master T
hesis.
Minn
esota: Uni
v
ersit
y
Of Minn
esota; 20
12.
[10]
G Hirst and
D
St-Onge.
Le
xic
a
l ch
ains
as re
prese
n
tatio
n
s
of conte
x
t for the d
e
tectio
n a
nd corr
ectio
n
of malapr
opis
m
s.
Christian
e
Fellb
au
m ed
isi
on, MIT Press.
199
8; 305-
332.
[1
1
]
JY Ji
an
g
,
SC Tsa
i
an
d SJ Le
e. FSKN
N
:
Mu
l
t
i-l
a
be
l
tex
t
ca
teg
o
r
i
z
a
t
i
o
n ba
sed
on
fu
zzy
si
mil
a
ri
ty
an
d k
near
est nei
ghb
ors.
Expert Systems w
i
th Appl
i
c
ations.
2
0
1
2
; 39:281
3-2
821.
[12]
J Slank
as an
d L W
ill
iams.
Automated E
x
traction of N
on-F
uncti
on
al
Req
u
ire
m
ents
in Avai
la
bl
e
Docu
mentatio
n
.
2013 1st Internatio
nal W
o
rksho
p
on N
a
tural L
a
n
gua
ge Ana
l
ysis i
n
Soft
w
a
r
e
Engi
neer
in
g (N
aturaL
iSE
)
. San F
r
ancisco, C
A
, USA, 2013; 9-16.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No
. 4, Decem
b
e
r
2015 : 145
6 – 1465
1465
[13]
Y Ko, S Park, J Seo and S Choi. Usi
ng cl
assi
ficati
on tec
hni
ques for inf
o
rmal req
u
ir
e
m
ents in th
e
requ
ireme
n
ts ana
l
y
sis-su
pp
o
r
ting
s
y
stem.
I
n
formatio
n
a
n
d
Softw
are T
e
chno
logy.
2
007
; 4
9
:
11
28
-
114
0.
[14]
D Li-g
uo, D
pe
ng a
nd L A
i
-p
i
ng. A Ne
w
Na
i
v
e Ba
yes T
e
xt
Classific
a
tio
n
Algorit
hm.
TELKOMNIKA.
201
4; 12: 947-
952.
[15]
GF
W
e
i Z
h
eng. F
eatur
e
Selectio
n
Method Bas
e
d on Improv
ed Doc
u
me
nt F
r
equenc
y.
TELKOMNIKA.
201
4; 12: 905-
910.
[16]
J Cle
lan
d
-Hu
a
ng, R S
e
ttimi, X Z
o
u a
nd P
Solc.
T
he
Dete
ction a
nd
Clas
s
ificatio
n of N
o
n-F
unctio
n
a
l
Req
u
ire
m
ents
w
i
th Applic
atio
n to Ear
l
y Asp
e
cts
. 14th IEE
E
Internati
ona
l
Confer
enc
e R
equ
ireme
n
ts
Engi
neer
in
g. Minn
eap
olis/St. Paul, MN. 200
6; 39-48.
[17]
K
Çelik an
d T
Güngör.
A
Co
mpre
he
nsi
v
e Ana
l
ysis o
f
using S
e
ma
ntic Informatio
n
in T
e
xt
Categori
z
a
tion
.
20
13
IEEE I
n
ternati
o
n
a
l S
y
mp
osi
u
m o
n
I
nnov
atio
ns i
n
Intel
lig
ent
S
y
stems
an
d
Appl
icatio
ns (INIST
A)
. Alben
a. 2013; 1-
5.
[18]
LS Jens
en
an
d T
Martinez.
Improv
in
g T
e
xt Classific
a
tio
n
by Us
ing
C
once
p
tua
l
an
d
Contextu
al
F
eatures
. Pro
c
eed
ings
of th
e W
o
rksh
op
o
n
T
e
xt Mi
nin
g
at the
6th
A
C
M SIGKDD I
n
ternati
o
n
a
l
Confer
ence
on
Kno
w
l
edg
e Di
scover
y
an
d D
a
ta Mini
ng. 20
00; 101-
10
2.
[19]
PROMISE,
"tera-PROMISE,"
1
April
201
0. [Online]
. Availa
b
l
e:
http://opensci
e
n
ce.us/rep
o
/re
quir
e
ments/oth
e
r-req
u
ir
em
ent
s/nfr.html. [Acc
essed 2
0
Jan
u
a
ri 20
15].
Evaluation Warning : The document was created with Spire.PDF for Python.