TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD
VS. CLASSICAL TF-IDF METHOD FOR WEB INFORMATION
EXTRACTION
Jo ge Rope o, A iel Gómez, Ca los León, Alejand o Ca asco
Depa men o Elec onic Technology,Uni e si y o Se ille, Se ille, Spain
Keywo ds: Te m weigh ing, TF-IDF, Fuzzy logic, In o ma ion ex ac ion, In o ma ion e ie al, Vec o space model,
In elligen agen .
Abs ac : Sol ing Te m Weigh ing p oblem is one o he mos impo an asks o In o ma ion Re ie al and
In o ma ion Ex ac ion. Tipically, he TF-IDF me hod ha e been widely used o de e mining he weigh o
a e m. In his pape , we p opose a no el al e na i e uzzy logic based me hod. The main ad an age o he
p oposed me hod is he ob en ion o be e esul s, especially in e ms o ex ac ing no only he mos
sui able in o ma ion bu also ela ed in o ma ion. This me hod will be used o he design o a Web
In elligen Agen which will soon s a o wo k o he Uni e si y o Se ille web page.
1 INTRODUCTION
The g ea amoun o a ailable in o ma ion caused by
he ising o In o ma ion Technology cons i u es an
eno mous ad an age when i comes o sea ch o
needed in o ma ion. Howe e , a he same ime, i is
a g ea p oblem o dis inguish he necessa y
in o ma ion among he huge quan i y o
unneccessa y da a.
Fo his eason, he concep s o In o ma ion
Re ie al (IR) and In o ma ion Ex ac ion (IE) came
up. IR is a ield in which he e ha e been g ea
ad ances in he las decades (Kwok, 1989),
especially in wha conce ns o he sea ch o
documen s. Ne e heless, IR does no only come
down o documen sea ching. IR ools may be used
o he objec s in any collec ion o accumula ed
knowledge such as he objec s s o ed in a shop o he
pho og aphies in an album. The gene aliza ion o
his me hod is possible hanks o he subs i u ion o
e e y objec o i s ep esen a ion in Na u al
Language (NL). IE in ol es a ans o ma ion o a
collec ion o documen s, gene ally helped by an IR
sys em. This collec ion o documen s is ans o med
in o easie o assimila e and analyze in o ma ion. IE
ies o ex ac ele an ac s om documen s,
whe eas IR selec s ele an documen s. The e o e, i
migh be said ha IE wo ks wi h a highe le el o
g anula i y han IR. (Kosala, 2002). In ou case, we
a e applying IE echniques o a web po al. A web
po al consis s o a collec ion o web pages, so he
me hod is comple ely applicable.
IR has been widely used o ex classi ica ion
(A onson e al., 1994; Liu e al., 2001) in oducing
app oaches such as Vec o Space Model (VSM), K
nea es neighbou me hod (KNN), Bayesian
classi ica ion model, neu al ne wo ks and Suppo
Vec o Machine (SVM) (Lu e al., 2002). VSM is
he mos equen ly used model. In VSM, a
documen is concep ually ep esen ed by a ec o o
keywo ds ex ac ed om he documen , wi h
associa ed weigh s ep esen ing he impo ance o
hese keywo ds in he documen . Typically, he so-
called TF-IDF me hod is used o de e mining he
weigh o a e m (Lee e al., 1997). Te m F equency
(TF) is he equency o occu ence o a e m in a
documen and In e se Documen F equency (IDF)
a ies in e sely wi h he numbe o documen s o
which he e m is assigned (Sal on & Buckley,
1988).
Al hough TF-IDF me hod o Te m Weigh ing
(TW) has wo ked easonably well o IR and has
been a s a ing poin o mo e ecen algo i hms,
(Lee e al., 1997; Sal on & Buckley, 1988; Liu e al.,
2001; Zhao & Ka ypis, 2002; Le na ee &
Thee amunkong, 2002; Xu e al., 2003), i was ne e
aken in o accoun ha some o he aspec s o
keywo ds may be impo an o de e mining e m
130
Rope o J., Gómez A., León C. and Ca asco A. (2009).
TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD VS. CLASSICAL TF-IDF METHOD FOR WEB INFORMATION EXTRACTION.
In P oceedings o he 11 h In e na ional Con e ence on En e p ise In o ma ion Sys ems - A i icial In elligence and Decision Suppo Sys ems, pages
130-137
DOI: 10.5220/0001982901300137
Copy igh c
SciTeP ess
weigh s apa om TF and IDF: i s o all, we
should conside he deg ee o iden i ica ion o an
objec i only he conside ed keywo d is used. This
pa ame e has a s ong in luence on he inal alue
o a e m weigh i he deg ee o iden i ica ion is
high. The mo e a keywo d iden i ies an objec , he
highe alue o he co esponding e m weigh ;
secondly, we should also conside he exis ance o
join e ms.
In his pape , we in oduce a uzzy logic (FL)
based e m weigh ing scheme. This scheme bea s in
mind all hese ea u es o calcula ing he weigh o
a e m, aking ad an age o uzzy logic lexibili y.
Fuzzy logic makes i possible o ha e a non- igid
e m weigh ing.
2 METHODOLOGY FOR IE
As said abo e, we a e applying IE o a web po al.
Pa icula ly, we ha e wo ked wi h he Uni e si y o
Se ille web po al. To ca y ou In o ma ion
Ex ac ion, i is necessa y o iden i y web page and
objec , ha is o say, e e y web page in a po al is
conside ed an objec . These objec s a e ga he ed in a
hie a chical s uc u e. An objec is classi ied unde a
unique c i e ion - o g oup o c i e ia -.
In ou case, we ha e aken ad an age o he
hie a chical s uc u e o a web page o di ide he
po al in h ee le els: Topic, Sec ion and Objec . The
size o e e y le el is a iable. E e y objec is
ep esen ed by means o a se o ques ions
o mula ed in NL. We ha e called hese ques ions
s anda d ques ions. The numbe o s anda d
ques ions associa ed o e e y web page is a iable,
depending on he amoun o in o ma ion con ained
in e e y page, i s impo ance and he numbe o
synonymous o index e ms. Logically, Sys em
Adminis a o ’s knowledge abou he ja gon o he
ela ed ield is p e y impo an . The highe his
knowledge, he highe eliabili y o he p oposed
s anda d ques ions, as hey shall be mo e simila o
possible use consul a ions. A e all, use s a e he
ones who y o ex ac he in o ma ion. Ou s udy
was based bo h on he s udy o he web pages
hemsel es and on p e ious consul a ions -
Uni e si y o Se ille bank o ques ions - .
Once s anda d ques ions a e de ined, index e ms
a e ex ac ed om hem. We ha e de ined hese
index e ms as wo ds, hough hey also may be
compound e ms. Index e ms a e he ones ha be e
ep esen a s anda d ques ion. E e y index e m is
associa ed wi h i s co esponden e m weigh . This
weigh has a alue be ween 0 and 1 and depends on
he impo ance o he e m in e e y hie a chic le el.
The highe impo ance in a le el, he highe is he
e m weigh . In addi ion, e m weigh is no cons an
o e e y le el, as he impo ance o a wo d o
dis inguish a opic om he o he s may be e y
di e en om i s impo ance o dis inguish be ween
wo objec s.
An example o he ollowed me hodology is
shown in Table 1.
Table 1: Example o he ollowed me hodology.
STEP EXAMPLE
S ep 1: Web page
iden i ied by s anda d/s
ques ion/s
- Web page:
www.us.es/uni i ual/in e ne
- S anda d ques ion : Which
se ices can I access as a i ual
use a he Uni e si y o Se ille?
S ep 2: Loca e
s anda d/s ques ion/s in
he hie a chical
s uc u e.
Topic 12: Vi ual Uni e si y
Sec ion 6: Vi ual Use
Objec 2.
S ep 3: Ex ac index
e ms
Index e ms: ‘se ices’, ‘ i ual’,
‘use ’
S ep 4: Te m weigh ing See sec ion 4
When a use consul a ion is made, hese e m
weigh s a e he inpu s o a uzzy logic sys em, which
mus de ec he objec o which he co esponden
use consul a ion e e s. Sys em ope a ion is
desc ibed in (Rope o e al., 2007).
3 TERM WEIGHTING
As said in p e ious sec ions o his pape , he e a e a
ew weigh s associa ed wi h e e y index e m. The
alues o he weigh s mus be ela ed somehow o
he impo ance o an index e m in i s co esponding
se o knowledge - in ou case, Topic, Sec ion o
Objec -. We may conside wo op ions o de ine
hese weigh s:
An expe in he ma e should e alua e
in ui i ely he impo ance o he index e ms. This
me hod is simple, bu i has he disad an age o
depending exclusi ely on he knowledge enginee . I
is e y subjec i e and i is no possible o au oma e
he me hod.
The gene a ion o au oma ed weigh s by means
o a se o ules. The mos widely used me hod o
TW is he TF-IDF me hod, bu we p opose a no el
Fuzzy Logic based me hod, which achie es be e
esul s in IE.
TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD VS. CLASSICAL TF-IDF METHOD FOR WEB
INFORMATION EXTRACTION
131
3.1 The TF-IDF Me hod
The idea o au oma ic ex e ie al sys ems based
on he iden i ica ion o ex con en and associa ed
iden i ie s is da ed in he 50s, bu i was Ge a d
Sal on in he la e 70s and he 80s who laid he
ounda ions o he exis ing ela ion be ween hese
iden i ie s and he ex s hey ep esen (Sal on &
Buckley, 1988). Sal on sugges ed ha e e y
documen D could be ep esen ed by e m ec o s k
and a se o weigh s wdk, which ep esen he weigh
o he e m k in documen D, ha is o say, i s
impo ance in he documen .
A TW sys em should imp o e e iciency in e ms
o wo main ac o s, ecall and p ecision. Recall
bea s in mind he ac ha he mos ele an objec s
o he use mus be e ie ed. P ecision akes in o
accoun ha s ange objec s mus be ejec ed. (Ruiz
& S ini asan, 1998). Recall may be de ined as he
numbe o e ie ed ele an objec s di ided by he
o al numbe o objec s. On he o he hand, p ecision
is he numbe o e ie ed ele an objec s di ided
by he o al numbe o e ie ed objec s. Recall
imp o es i high- equency e ms a e used, as such
e ms will make i possible o e ie e many objec s,
including he ele an ones. P ecision imp o es i
low- equency e ms a e used, as speci ic e ms will
isola e he ele an objec s om he non- ele an
ones. In p ac ice, comp omise solu ions a e used,
using e ms which a e equen enough o each a
easonable le el o ecall wi hou p oducing a oo
low p ecision.
The e o e, e ms ha a e men ioned o en in
indi idual objec s, seem o be use ul o imp o e
ecall. This sugges s he u iliza ion o a ac o named
Te m F equency (TF). Te m F equency (TF) is he
equency o occu ence o a e m. On he o he side,
ano he ac o should a o he e ms concen a ed in
a ew documen s o he collec ion. The in e se
equency o documen (IDF) a ies in e sely wi h
he numbe o objec s (n) o which he e m is
assigned in an N-objec collec ion. A ypical IDF
ac o is log (N/n). (Sal on & Buckley, 1988). A
usual o mula o desc ibe he weigh o a e m j in
documen i is:
wi
j
= i
j
x id
j
. (1)
This o mula has been modi ied and imp o ed by
many au ho s o achie e be e esul s in IR and IE
(Lee e al., 1997; Liu e al., 2001; Zhao & Ka ypis,
2002; Le na ee & Thee amunkong, 2002; Xu e al.,
2003).
3.2 The FL based Me hod
The TF-IDF me hod wo ks easonably well, bu i
has he disad an age o no conside ing wo key
aspec s o us:
The deg ee o iden i ica ion o he objec i only
he conside ed index e m is used. This pa ame e
has a s ong in luence on he inal alue o a e m
weigh i he deg ee o iden i ica ion is high. The
mo e a keywo d iden i ies an objec , he highe
alue o he co esponding e m weigh .
Ne e heless, his pa ame e c ea es wo
disad an ages in e ms o p ac ical aspec s when i
comes o ca ying ou a e m weigh au oma ed and
sys ema ic assignmen . On he one hand, he deg ee
o iden i ica ion is no deduc ible om any
cha ac e is ic o a keywo d, so i mus be speci ied
by he Sys em Adminis a o . On he second hand,
he same keywo d may ha e a di e en ela ionship
wi h e e y objec .
The second pa ame e is ela ed o join e ms. In
he index e m ‘ e m weigh ing’, his exp ession
would cons i u e a join e m. E e y single e m in a
join e m has a lowe alue han i would ha e i i
did no belong o i . Howe e , i we combine all he
single e ms in a join e m, e m weigh mus be
highe . A join e m may eally de e mine an objec
whe eas he appea ance o only one o i s single
e ms may e e o ano he objec .
The conside a ion o hese wo pa ame e s
oge he wi h classical TF and IDF de e mines he
weigh o an index e m o e e y subse in e e y
le el. The FL based me hod gi es a solu ion o all
he p oblems and also gi es wo main ad an ages.
The solu ion o bo h p oblems is o c ea e a able
wi h all he keywo ds and hei co esponding
weigh s o e e y objec . This able will be c ea ed
in he phase o keywo d ex ac ion om s anda d
ques ions. Imp ecision p ac ically does no a ec he
wo king me hod due o he ac ha bo h e m
weigh ing and in o ma ion ex ac ion a e based on
uzzy logic, wha minimizes possible a ia ions o
he assigned weigh s. The way o ex ac ing
in o ma ion also helps o success ully o e come his
imp ecision. In addi ion, he FL based me hod also
gi es impo an ad an ages: on he one hand, e m
weigh ing is au oma ed; on he o he hand, he le el
o equi ed expe ise o an ope a o is lowe . This
ope a o would no need o know any hing abou he
FL engine unc ioning, bu only how many imes
does a e m appea in any subse and he answe o
hese ques ions: a) Does a keywo d undoub edly
ICEIS 2009 - In e na ional Con e ence on En e p ise In o ma ion Sys ems
132
de ine an objec by i sel ? b) Is a keywo d ied o
ano he one?
In ou case, he applica ion o his me hod o a
web po al, he web po al de elope himsel may
de ine simul aneously he s anda d ques ions and
index e ms associa ed wi h he objec - a web page -
and he esponse o he ques ions men ioned abo e.
4 METHOD IMPLEMENTATION
This sec ion shows how he TF-IDF me hod and he
FL based me hod we e implemen ed in p ac ise, in
o de o compa e bo h me hods applying hem o he
Uni e si y o Se ille web po al.
4.1 TF-IDF Me hod Implemen a ion
As men ioned in p e ious sec ions, a easonable
measu e o he impo ance o a e m may be
ob ained by means o he TF-IDF p oduc . Howe e ,
his o mula has been modi ied and imp o ed by
many au ho s o achie e be e esul s in IR and IE.
E en ually, he chosen o mula o ou es s was he
one p oposed by Liu e al. (Liu e al., 2001).
∑
=
+×
+×
=
m
k
kik
kik
ik
nN
nN
W
1
2
))01.0/log(
)01.0/log(
(2)
Whe e ik is he i h e m equency o
occu ence in he k h subse - Topic / Sec ion /
Objec -. nk is he numbe o subse s o which he
e mTk is assigned in a collec ion o N objec s.
Consequen ly, i is aken in o accoun ha a e m
migh be p esen in o he se s o he collec ion.
As an example, we a e using he e m ‘ i ual’,
abo e used in he example in Sec ion 2.
A Topic le el:
- ‘Vi ual’ appea s 8 imes in Topic 12 ( ik = 8,
K=12).
- ‘Vi ual’ appea s wice in o he Topics (nk = 3)
- The e a e 12 Topics in o al (N=12) - o
no malizing, i is only necessa y o know he
o he ik and nk o he Topic -.
- Subs i u ing, Wik = 0.20.
A Sec ion le el:
- ‘Vi ual’ appea s 3 imes in Sec ion 12.6 ( ik =
3, K= 6)
- ‘Vi ual’ appea s 5 imes in o he Sec ions in
Topic 12 (nk = 6)
- The e a e 6 Sec ions in Topic 12 (N=6).
- Subs i u ing, Wik = 0.17.
A Objec le el:
- ‘Vi ual’ appea s once in Objec 12.6.2 ( ik = 1,
K = 2). – Logically a e m can only appea once
in an Objec -.
- ‘Vi ual’ appea s wice in o he Topics (nk = 3)
- The e a e 3 Objec s in Sec ion 12.6 (N=3).
- Subs i u ing, Wik = 0.01. In ac , ‘ i ual’
appea s in all he Objec s in Sec ion 12.6, so i is
i ele an o dis inguish he Objec .
Consequen ly, ‘ i ual’ will be ele an o ind
ou ha he Objec is in Topic 12, Sec ion 6, bu
i ele an o ind ou he de ini e Objec , which
should be ound acco ding o o he e ms in a use
consul a ion.
4.2 FL based Me hod Implemen a ion
As said in sec ion 3.2, TF-IDF has he disad an age
o no conside ing he deg ee o iden i ica ion o he
objec i only he conside ed index e m is used and
he exis ance o ied keywo ds. Like TF-IDF
me hod, i is neccesa y o know TF and IDF, and
also he answe o he ques ions men ioned in
sec ion 3.2. FL based Te m Weigh ing me hod is
de ined below. Fou ques ions mus be answe ed o
de e mine he Te m Weigh o an Index Te m:
- Ques ion 1 (Q1): How o en does an index e m
appea in o he subse s? - Rela ed o IDF -.
- Ques ion 2 (Q2): How o en does an index e m
appea in i s own subse ? - Rela ed o TF -.
- Ques ion 3 (Q3): Does an index e m
undoub edly de ine an objec by i sel ?
- Ques ion 4 (Q4): Is an index e m ied o
ano he one?
Ques ion 1
Te m weigh is pa ly associa ed o he ques ion
‘How o en does an index e m appea in o he
subse s?’. I is gi en by a alue be ween 0 – i i
appea s many imes – and 1 - i i does no appea in
any o he subse -. To de ine weigh s, we a e
conside ing he imes ha he mos used e ms in he
whole se o knowledge appea .
P o ided ha he e a e 1114 index e ms de ined
in ou case, we ha e assumed ha 1 % o hese
wo ds mus ma k he bo de o he alue 0 (11
wo ds). As he ele en h mos used wo d appea s 12
imes, whene e an index e m appea s mo e han 12
imes in o he subse s, we will gi e i he alue o 0.
Values o e e y Topic a e de ined in Table 2.
TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD VS. CLASSICAL TF-IDF METHOD FOR WEB
INFORMATION EXTRACTION
133
Table 2: Te m weigh alues o e e y Topic o Q1.
Times
appea ing
0 1 2 3 4 5 6 7 8 9 10 11 12 >12
Value 1 0.9 0.8 0.7 0.64 0.59 0.53 0.47 0.41 0.36 0.3 0.2 0.1 0
Table 3: Te m weigh alues o e e y Sec ion o Q1.
Times appea ing 0 1 2 3 4 5 > 5
Value 1 0.7 0.6 0.5 0.4 0.3 0
Table 4: Te m weigh alues o e e y Objec o Q1.
Times appea ing 0 1 2 >2
Value 1 0.7 0.3 0
Table 5: Te m weigh alues o e e y Topic and Sec ion o Q2.
Times appea ing 1 2 3 4 5 > 5
Value 0 0.3 0.45 0.6 0.7 1
Table 6: Te m weigh alues o Q3.
Answe o Q3: Does a e m de ine undoub edly
a s anda d ques ion?
Yes Ra he No
Value 1 0.5 0
Be ween 0 and 3 imes appea ing -
app oxima ely a hi d o he possible alues - , we
conside ha an index e m belongs o he so called
HIGH se . The e o e, i is de ined in i s
co espondan uzzy se wi h uni o mly dis ibu ed
alues be ween 0.7 and 1, as may be seen in Figu e
1. Analogously, we may dis ibu e all alues
uni o mly acco ding o di e en uzzy se s. Fuzzy
se s a e iangula , on one hand o simplici y and on
he o he hand because we es ed o he mo e
complex ypes o se s (Gauss, Pi ype, e c) and he
esul s did no imp o e a all.
Figu e 1: Inpu uzzy se s.
P o ided ha di e en weigh s a e de ined in
e e y hie a chic le el, we should conside o he
scales o calcula e hem. As o he Topic Le el we
we e conside ing he immedia ely op le el – he
whole se o knowledge - , o he Sec ion le el we
should conside he imes ha an index e m appea s
in a ce ain Topic. We again conside ha 1 % o
hese wo ds mus ma k he bo de o he alue 0 -
11 wo ds - . The ele en h mos used index e m in a
unique Topic appea s 5 imes, so whene e a e m
appea s mo e han 5 imes in o he subse s, i s
weigh akes he alue 0 a he Sec ion o le el.
Possible e m weigh s o he le el o Sec ion a e
shown in Table 3. The me hod is analogous and
conside s he de ini ion o he uzzy se s. A he
le el o Objec , e m weigh s a e shown in Table 4.
Ques ion 2
To ind ou he e m weigh associa ed o ques ion 2
- Q2, How o en does an index e m appea in i s
own subse ? -, he easoning is analogous. Howe e ,
we ha e o bea in mind ha i is necessa y o
conside he equency inside a unique se o
knowledge, hus he numbe o appea ances o index
e ms dec eases conside ably. The lis o he mos
used index e ms in a Topic mus be conside ed
again. I also mus be bo n in mind ha he mo e an
index e m appea s in a Topic o Sec ion, he highe
alue o an index e m. Q2 is senseless a he le el
o Objec . The p oposed alues a e gi en in Table 5.
ICEIS 2009 - In e na ional Con e ence on En e p ise In o ma ion Sys ems
134

Table 7: Te m weigh alues o Q4.
Numbe o index e ms ied o
ano he index e m
0 1 2 > 2
Value 1 0.7 0.3 0
Table 8: Rule de ini ion o Topic and Sec ion le els.
Rule numbe Rule de ini ion Ou pu
R1 IF Q1 = HIGH and Q2 ≠ LOW A leas MEDIUM-HIGH
R2 IF Q1 = MEDIUM and Q2 = HIGH A leas MEDIUM-HIGH
R3 IF Q1 = HIGH and Q2 = LOW Depends on o he Ques ions
R4 IF Q1 = HIGH and Q2 = LOW Depends on o he Ques ions
R5 IF Q3 = HIGH A leas MEDIUM-HIGH
R6 IF Q4 = LOW Descends a le el
R7 IF Q4 = MEDIUM I he Ou pu is MEDIUM-
LOW, i descends o LOW
R8 IF (R1 and R2) o (R1 and R5) o (R2 and R5) HIGH
R9 In any o he case MEDIUM-LOW
Ques ion 3
In he case o ques ion 3 – Q3, Does a e m de ine
undoub edly a s anda d ques ion? - , he answe is
comple ely subjec i e and we p opose he answe s
‘Yes’, ‘Ra he ’ and ‘No’. Te m weigh alues o
his ques ion a e shown in Table 6.
Ques ion 4
Finally, ques ion 4 – Q4, Is an index e m ied o
ano he one? – deals wi h he numbe o index e ms
ied o ano he one. We p opose e m weigh alues
o his ques ion in Table 7. Again, he alues 0.7
and 0.3 a e a consequence o conside ing he bo de
be ween uzzy se s – see Figu e 1-.
A e conside ing all hese ac o s, uzzy ules
o Topic and Sec ion le els a e de ined in Table 8.
This ules co e all he 81 possible combina ions.
No e ha , apa om he h ee inpu se s men ioned
in p e ious sec ions, ou ou pu se s ha e been
de ined - HIGH, MEDIUM-HIGH, MEDIUM-LOW
and LOW-, as may be seen in Figu e 2. A he le el
o Objec , we mus disca d ques ion 2 and ules
change.
The only aspec which has no been de ined ye
is abou mul iple appea ances in a Topic o Sec ion.
I.e., i is possible ha he answe o ques ion 3 is
‘Ra he ’ in one case ‘No’ in ano he one. In his
case, a weigh ed a e age o he co esponding e m
weigh s is calcula ed.
An example o all he p ocess is shown below
Example
Objec 12.6.2 is de ined by he ollowing s anda d
ques ion :
Which se ices can I access as a i ual use
a he Uni e si y o Se ille?
I we conside he e m ‘ i ual’:
- A Topic le el:
- ‘Vi ual’ appea s wice in o he Topics in he
whole se o knowledge, so ha he alue associa ed
o Q1 is 0.80.
Figu e 2: Ou pu uzzy se s.
- ‘Vi ual’ appea s 8 imes in Topic 12, so ha
he alue associa ed o Q2 is 1.
- The esponse o Q3 is ‘Ra he ’ in 5 o he 8
imes and ‘No’ in he o he h ee, so ha he alue
associa ed o Q3 is a weigh ed a e age: (5*0.5 +
3*0)/8 = 0.375.
- Te m ‘ i ual’ is ied o one e m 7 imes and i
is ied o wo e ms once. The e o e, he a e age is
TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD VS. CLASSICAL TF-IDF METHOD FOR WEB
INFORMATION EXTRACTION
135
Table 9: Tes compa ison be ween bo h me hods.
Ca 1 Ca 2 Ca 3 Ca 4 Ca 5 To al
TF-IDF
Me hod
466
(50.98%)
223
(24.40%)
53
(5.80%)
79
(8.64%)
93
(10.18%)
914
FL Me hod 710
(77.68%)
108
(11.82%)
27
(2.95%)
28
(3.06%)
41
(4.49%)
914
1.14 e ms. A linea ex apola ion leads o a alue
associa ed o Q4 o 0.65.
- Wi h all he alues as inpu s o he uzzy
logic engine, we ob ain a e m weigh o 0.53.
- A Sec ion le el
- ‘Vi ual’ appea s 5 imes in o he Sec ions
co esponding o Topic 12, so ha he alue
associa ed o Q1 is 0.30.
- ‘Vi ual’ appea s 3 imes in Topic 12, so ha
he alue associa ed o Q2 is 0.45.
- The esponse o Q3 is ‘Ra he ’ in all cases, so
ha he alue associa ed o Q3 is 0.5.
- Te m ‘ i ual’ is ied o e m ‘use ’ so ha he
alue associa ed o Q4 is 0.7.
- Wi h all he alues as inpu s o he uzzy
logic engine, we ob ain a e m weigh o 0.45.
- A Objec le el:
- ‘Vi ual’ appea s wice in o he Objec s
co esponding o Sec ion 12.6, so ha he alue
associa ed o Q1 is 0.30.
- The esponse o Q3 is ‘Ra he ’, so ha he
alue associa ed o Q3 is 0.5.
- Te m ‘ i ual’ is ied o e m ‘use ’ so ha he
alue associa ed o Q4 is 0.7.
- Wi h all he alues as inpu s o he uzzy
logic engine, we ob ain a e m weigh o 0.52. We
can see he di e ence wi h he co esponding e m
weigh ob ained wi h he TF-IDF me hod, bu i is
exac ly wha we a e looking o : no only he
desi ed objec mus be e ie ed, bu he mos
closely ela ed o i .
5 TESTS AND RESULTS
Tes s ha e been done on he Uni e si y o Se ille
web po al. This web po al has 50,000 daily isi s,
wha quali ies i in o he 10% mos isi ed
Uni e si y po als – he e a e mo e han 4,000 -.
As he e is much in o ma ion in i , 253 objec s
g ouped in 12 Topics we e de ined. All hese
g oups we e made up o a a iable numbe o
Sec ions and Objec s. 2107 s anda d ques ions
su ged om hese 253 Objec s, bu sligh ly mo e
han he hal o hem we e elimina ed o hese
es s because o being e y simila o o he s.
E en ually, es s consis ed o 914 use
consul a ions.
To compa e esul s, we conside ed he posi ion
in which he co ec answe appea ed among he
e ie ed answe s, acco ding o uzzy engine
ou pu s. The i s necessa y s ep o ollow is o
de ine he o e coming h esholds o he uzzy
engine. This way, Topics and Sec ions ha a e no
ela ed wi h he Objec o iden i y a e elimina ed.
We also ha e o de ine low enough h esholds, in
o de o be able o ob ain also ela ed Objec s. We
sugges o p esen be ween 1 and 5 answe s,
depending on he numbe o ela ed Objec s.
The esul s o he consul a ion we e so ed in 5
ca ego ies:
- Ca ego y Ca 1: he co ec answe is e ie ed
as he only answe o i is he one ha has a highe
deg ee o ce ain y be ween he answe s e ie ed
by he sys em.
- Ca ego y Ca 2: The co ec answe is
e ie ed be ween he 3 wi h highe deg ee o
ce ain y -excluding he p e ious case -.
- Ca ego y Ca 3: The co ec answe is
e ie ed be ween he 5 wi h highe deg ee o
ce ain y - excluding he p e ious cases -.
- Ca ego y Ca 4: The co ec answe is
e ie ed, bu no be ween he 5 wi h highe deg ee
o ce ain y.
- Ca ego y Ca 5: The co ec answe is no
e ie ed by sys em.
The ideal si ua ion comes when he desi ed
Objec is e ie ed as Ca 1, hough Ca 2 and Ca 3
would be easonably accep able. The ob ained
esul s a e shown in Table 9. Though he ob ained
esul s wi h he TF-IDF me hod a e qui e
easonable, 81.18 % o he objec s being e ie ed
be ween he i s 5 op ions - and mo e han as
Ca 1, he FL based me hod u ns ou o be clea ly
be e , wi h 92.45 % o he desi ed Objec s
e ie ed - and mo e han h ee qua e s as he i s
op ion -.
ICEIS 2009 - In e na ional Con e ence on En e p ise In o ma ion Sys ems
136
6 CONCLUSIONS
A FL based Te m Weigh ing me hod has been
p esen ed as an al e na i e o classical TF-IDF
me hod.
The main ad an age o he p oposed me hod is
he ob en ion o be e esul s, especially in e ms
o ex ac ing no only he mos sui able
in o ma ion bu also ela ed in o ma ion.
This me hod will be used o he design o a
Web In elligen Agen which will soon s a o
wo k o he Uni e si y o Se ille web page.
REFERENCES
A onson, A.R, Rind lesch, T.C, B owne, A. C., 1994.
Exploi ing a la ge hesau us o in o ma ion
e ie al. P oceedings o RIAO, pp. 197-216.
Kosala, R., Blockeel, H., 2002. Web Mining Resea ch: A
Su ey. SIGKDD Explo a ions: Newsle e o he
Special In e es G oup (SIG) on Knowledge
Disco e y & Da a Mining, ACM, Vol. 2 (2000).
Kwok, K. L., 1989. A neu al ne wo k o p obabilis ic
in o ma ion e ie al. P oceedings o he 12 h annual
in e na ional ACM SIGIR con e ence on Resea ch
and de elopmen in in o ma ion e ie al.
Camb idge, Massachuse s, Uni ed S a es.
Lee, D.L., Chuang, H., Seamons, K., 1997. Documen
anking and he ec o -space model. IEEE So wa e,
Vol. 14, Issue 2, p. 67 – 75.
Le na ee, V., Thee amunkong, T. 2003. Combining
homogenous classi ie s o cen oid-based ex
classi ica ion. P oceedings o he 7 h In e na ional
Symposium on Compu e s and Communica ions, pp.
1034-1039.
Liu, S., Dong, M., Zhang, H., Li, R. Shi, Z., 2001. An
app oach o mul i-hie a chy ex classi ica ion
P oceedings o he In e na ional Con e ences on In o-
ech and In o-ne , 2001. Beijing. Vol 3, pp. 95 – 100.
Lu, M., Hu, K., Wu, Y., Lu, Y., Zhou, L., 2002.
SECTCS: owa ds imp o ing VSM and Nai e
Bayesian classi ie . IEEE In e na ional Con e ence
on Sys ems, Man and Cybe ne ics, Vol. 5, p. 5.
Rope o J., Gomez, A., Leon, C., Ca asco, A. 2007.
In o ma ion Ex ac ion in a Se o Knowledge Using
a Fuzzy Logic Based In elligen Agen . Lec u e
No es in Compu e Science. Vol. 4707, pp. 811-820.
Ruiz, M.E., S ini asan, P., 1998. Au oma ic Tex
Ca ego iza ion Using Neu al Ne wo ks. Ad ances in
Classi ica ion Resea ch ol. 8: P oceedings o he 8 h
ASIS SIG/CR Classi ica ion Resea ch Wo kshop. Ed.
E himis E himiadis. In o ma ion Today,
Med o d:New Je sey. 1998. pp 59-72.
Sal on, G., Buckley, C., 1996. Te m Weigh ing
App oaches in Au oma ic Tex Re ie al. In o ma ion
P ocessing and Managemen , Vol.32 (4), pp. 431-
443.
Xu, J., Wang, Z. 2003. TCBLSA: A new me hod o ex
clus e ing. In e na ional Con e ence on Machine
Lea ning and Cybe ne ics. Vol. 1, pp. 63-66.
Zhao, Y., Ka ypis, G., 2002. Imp o ing p eca ego ized
collec ion e ie al by using supe ised e m
weigh ing schemes. P oceedings o he In e na ional
Con e ence on In o ma ion Technology: Coding and
Compu ing, 2002. pp 16 – 21.
TERM WEIGHTING: NOVEL FUZZY LOGIC BASED METHOD VS. CLASSICAL TF-IDF METHOD FOR WEB
INFORMATION EXTRACTION
137