A Me hod o he Access o he Con en s in a Se o Knowledge Using a Fuzzy
Logic Based In elligen Agen
Jo ge Rope o1, A iel Gómez1, Ca los León1 and Alejand o Ca asco1
1 Depa men o Elec onic Technology, Uni e si y o Se ille, Spain.
j ope o@d e.us.es, a [email protected], [email protected], aca [email protected].
Escuela Técnica Supe io de Ingenie ía In o má ica
A da. Reina Me cedes, s/n
41012 - Se illa (Spain)
Abs ac
This pape p oposes a me hod o he classi ica ion o
he con en s in a se o knowledge in o de o answe o
use consul a ions using na u al language. The sys em is
based on a uzzy logic engine, which akes ad an age o i s
lexibili y o managing se s o accumula ed knowledge.
These se s can be buil in hie a chic le els by a ee
s uc u e. A me hod o consul a ion based on a uzzy logic
applica ion p o ided wi h an in e ace ha one may
in e ac wi h in na u al language is also p oposed. The
e en ual aim o his sys em is he implemen a ion o an
in elligen agen o manage he in o ma ion con ained in
an in e ne po al.
1. Mo i a ions
The access o he con en s o an ex ensi e se o
accumula ed knowledge – a da abase, a summa y o
documen s, web con en s, pic u es, e c – is an impo an
conce n nowadays. In o ma ion Re ie al (IR) deals wi h
la ge collec ions o ex ual ma e ial, and i s aim is o
sa is y use que ies and needs [1]. These needs a e
inc eased when he use in ques ion is no amilia wi h he
ma e o he e a e ambiguous con en s, bad o ganiza ion
o , simply, complex opics o a g ea amoun o
in o ma ion di icul o manage.
E en ually, unsuccess ul a emp s can u n ou o be
us a ing i he exac e m o e ms a e no used o make
he consul a ions - a machine only will answe adequa ely
i i is asked in an exac way -, and one can e en ually end
in a pa adox: he less one knows he mo e di icul i is o
ind he answe s. In many cases he solu ion is o seek help
om an expe on he opic. In ac he pe son asked o
help is an in e p e e who is able o gene a e a syn ac ically
and seman ically a co ec sea ch ob aining he desi ed
answe s. Consequen ly, he e is he need o an agen o
in e p e he ague in o ma ion we p o ide, gi ing us
conc e e answe s ela ed o he exis ing con en s o he se
o knowledge. This should be based on an es ima ion o
he ce ain y o he ela ion be ween wha we ha e
exp essed in na u al language and he con en s s o ed in
he se o knowledge.
To sol e his, we ha e de eloped a me hod o
classi ica ion o con en s by c ea ing a ew indexes based
on key wo ds, and a me hod o consul a ion based on a
uzzy logic applica ion wi h an in e ace ha one may
in e ac wi h in na u al language. We hen p opose an
a i icial in elligence (AI) applica ion based on he use o
uzzy logic.
2. Mode o ope a ion
All p in ed ma e ial, including ex , illus a ions, and
cha s, mus be kep wi hin a p in a ea o 6.9 inches (175
mm) wide by 8.9 inches (226 mm) high. Do no w i e o
p in any hing ou side he p in a ea. The op ma gin mus
be 1 inch (25 mm), excep o he i le page, and he le
ma gin mus be 0.75 inch (19 mm). All ex mus be in a
wo-column o ma . Columns a e o be 3.27 inches (83
mm) wide, wi h a 0.37 inch (8 mm) space be ween hem.
Tex mus be ully jus i ied.
2.1. Objec i es
The main objec i e o he sys em designed mus be o
le he use s ind possible answe s o wha hey a e looking
o in a huge se o knowledge. Wi h his aim, he whole
se o knowledge mus be classi ied in o di e en objec s,
as shown in Figu e 1. These objec s a e he possible use
consul a ions, o ganized in hie a chic g oups. A s anda d
ques ion is assigned o e e y objec , and di e en key
wo ds om each s anda d ques ion mus hen be selec ed
in o de o di e en ia e one objec om he o he s. Finally,
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007
Figu e 1. Gene a ed s anda d ques ion in e ace.
Figu e 2. Te m weigh ing scheme.
e m weigh s a e assigned o e e y wo d o e e y le el o
hie a chy in a scheme based on a ec o space model.
These 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.
2.2. Hie a chic s uc u e
The s anda d-ques ions collec ion gene a ed om he
whole se o knowledge mus be o ganized in hie a chic
g oups. This o e s he ad an age o easie handling. Fo
es s, we decided o use he ques ions-answe s da abase o
he Uni e si y o Se ille o c ea e a lis o he mos
equen ly asked ques ions. The whole knowledge was
o ganized in o h ee le els: opic, sec ion and ques ion. Fo
any web page, i would be enough o c ea e a bank o
possible ques ions–answe s, a ange hem hie a chically,
and iden i y hem as objec s.
E e y possible use ques ion mus be ela ed wi h
s anda d ques ions in o de o p esen i s s anda d answe s
as possible answe s o his use ques ion. This may be seen
in Figu e 3. The e a e ob iously many ways o asking, so
he aim o he sys em is o iden i y eal use consul a ions
and wha we ha e called s anda d ques ions. As he e may
be se e al ques ions simila o he one asked by he use o
as i may be in e es ing o he use o ge ela ed answe s,
se e al answe s will be p esen ed.
When a ques ion is made, he i s p ocessing s ep
consis s o dis inguishing he key wo ds in he
consul a ion. These wo ds mus be sea ched in a da abase
which mus con ain all he wo ds ha a e ela ed somehow
o he con en o he subjec we a e dealing wi h –see
Figu e 4. Ano he da abase wi h he possible answe s
becomes necessa y.
Key wo ds a e assigned o e e y s anda d ques ion in
o de o iden i y i . These wo ds a e chosen among hose
ha could appea in a possible consul a ion. As men ioned
abo e, he whole knowledge is g ouped in a ious
hie a chic le els, so he belonging o hese wo ds o
di e en le els is de e mined by a ew nume ical
coe icien s indica ing how signi ican he conside ed wo d
is wi hin he le el in ques ion. Wi h his aim, weigh
ec o s a e assigned o each wo d. Each ec o con ains he
ce ain y o belonging o e e y uzzy se . I is impo an o
no ice ha he same wo d can belong o se e al di e en
se s.
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007
Figu e 3. Mode o ope a ion.
Figu e 4. Index wo ds da abase.
2.3. Te m weigh ing
As he sys em mus wo k using a na u al language
in e ace, we p opose he use o a sys em based on ec o
space modeling o e m weigh ing. Au oma ic documen
indexing is usually based on he equency o occu ence,
ha is o say, wo ds wi h a high equency o occu ence
a e less signi ican han wo ds wi h a low one, and he
in e se documen equency, ha is, how o en i occu s in
he whole documen . The main limi a ion o his scheme is
c ea ed by e ms which a e unhelp ul o iden i ica ion bu
ha e he same equency o occu ence as he ele an ones
[2-8].
Thus, we p opose a no el scheme based on a ec o
space model which also akes in o accoun i he wo d is
impo an o he meaning o he ques ion i sel o i he
wo d is linked o o he wo ds and i s ela ionship wi h he
whole se o knowledge o e m weigh ing [7,9].
The success o he p oposed me hod depends o a g ea
ex en on a co ec assignmen o he coe icien s o he
key wo ds, ha is, he building o he weigh ec o s. The
p ocess consis s o 3 s ages: elec ion o key wo ds,
coe icien assignmen o he chosen wo ds; and
modi ica ion o he index alues in o de o ob ain he
desi ed minimal ce ain ies.
2.3.1. Elec ion o key wo ds. As men ioned ea lie , o
e e y elemen ha has o be de e mined, a ques ion is
assigned in na u al language. These ques ions a e called
s anda d ques ions. F om e e y s anda d ques ion key
wo ds a e chosen. These key wo ds will allow he use
consul a ion and hese s anda d ques ions o be ma ched
and, he e o e, o indica e he equi ed elemen . Wo d
elec ion is based on i s conc e ion, meaning he deg ee o
ela ion o he wo d wi h he elemen o be iden i ied. This
excludes a icles, conjunc ions, e b o ms, e c, unless hey
a e s ongly signi ican o he ques ion s uc u e.
2.3.2. Coe icien assignmen . Fo e e y key wo d,
coe icien s co esponding o e e y le el o hie a chy o
he whole se o knowledge mus be assigned. The highe
he ela ionship be ween he p esen le el and he key
wo d, he highe he coe icien o ha key wo d will be
o ha le el. These coe icien s a e he inpu s o e e y
uzzy logic sys em, as desc ibed in Sec ion 2.4.
2.3.3. S anda d ques ion ecogni ion es . Once he
coe icien assignmen is made, s anda d ques ion
ecogni ion es s mus ake place. These es s a e based on
using s anda d ques ions as use consul a ions.
2.3.4. Coe icien modi ica ion. The esul s ob ained o
s anda d ques ion ecogni ion es s ep esen he
ela ionship o e e y s anda d-ques ion ela ionship wi h
i sel . The highe i is he be e he esul s a e. Howe e , i
mus be aken in o accoun ha simila ques ions mus
ha e a high deg ee o ce ain y oo. This allows o he
modi ica ion o coe icien s in o de o achie e be e
esul s. An example o how we ha e modi ied some
coe icien s is shown in Sec ion 3.1.
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007
2.4. Fuzzy logic sys em
Fuzzy logic a ises as a esponse o he in lexibili y o
he classic bina y logic [10-11]. By means o a se o
unc ions, a deg ee o lexibili y may be gi en o hese
epi he s: wha may be cool o a Se illian migh be mild
o a Be line .
A uzzy logic sys em gi es lexibili y o e m
weigh ing. Mo e impo an han ha ing a conc e e alue
o weigh s, wha eally ma e s is ha a ea u e is
ep esen ed by a wo d. I is no so impo an ha a weigh
is 0.8 o 0.9: he weigh is HIGH in bo h cases.
All he key wo ds a e ex ac ed o compa ison wi h
he ones con ained in ou key wo d da abase. As
men ioned ea lie , he whole se o knowledge is a anged
in le els o hie a chy. The inpu s o he uzzy logic sys em
a e he coe icien s o belonging o e e y le el. The se
“belonging o e e y le el 1” is analyzed bea ing in mind
he alue e u ned by he uzzy engine. I he le el o
ce ain y is lowe han a p ede ined alue, he con en o
he co esponding se is ejec ed. S a ing om le el one
and using a ee s uc u e makes i possible o ejec a g ea
amoun o con en which will no be conside ed in u u e
sea ches.
Fo e e y se ha has o e come a minimum ce ain y
h eshold, he p ocess is epea ed and he coe icien s o
belonging co esponding o e e y le el 2 se a e e alua ed.
Se s whe e he deg ee o ce ain y e u ned by he uzzy
engine does no su pass a ce ain minimal h eshold a e
ejec ed. I hey su pass he h eshold, he me hod o
de e mining he belonging o he ollowing le el is applied
o hem. This p ocess is epea ed un il he las le el. The
answe s co espond o hose las le el elemen s in which
ce ain y has o e come he de ini e h eshold. The e can be
mo e han one answe . The ague he ques ions, he mo e
answe s we will ob ain. This p ocess may be seen in
Figu e 5.
Figu e 5. Fuzzy logic sys em: ope a ion mode.
The hea o he uzzy logic sys em is he uzzy engine.
This engine is esponsible o de e mining he p obabili y
ha he key wo ds con ained in a consul a ion will belong
o a ce ain uzzy se in a speci ic le el. The engine mus
e alua e he belonging o e e y se o he co esponding
le el. Thus, he engine akes he coe icien s o he key
wo ds o ha se as inpu s. The uzzy engine ou pu will
be de e mined by he de ined ules. These ules a e o he
IF … THEN ype. An example o a ule migh be his:
IF wo d_index1 is HIGH AND wo d_index2 is MEDIUM
AND wo d_index3 is LOW, THEN ou pu is HIGH.
2.5. Sys em adminis a o
The sys em equi es a sys em adminis a o who has
basically h ee unc ions:
a) De ining and modi ying e m weigh s and ules.
b) Adding new wo ds o he da abase when necessa y.
c) C ea ing a sys em eedback which asks he e en ual
use s o hei opinion abou he answe s gi en by he
assis an in o de o ake he necessa y s eps in each case.
Thus a , an adminis a o in e ace has been c ea ed –
Figu e 6 - .
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007
Figu e 6. Adminis a o in e ace.
3. Tes s and esul s
3.1. Tes de ini ion
Tes s a e based on he use o s anda d ques ions as use
consul a ions. The i s goal o hese es s is o check ha
he sys em makes a co ec iden i ica ion o s anda d
ques ions wi h an index o ce ain y highe han 0.7. The
use o uzzy logic makes i possible o iden i y no only he
co esponding s anda d ques ion bu o he s as well. This is
ela ed o ecall, hough i does no ma ch ha exac
de ini ion [12]. The second goal is o check i he equi ed
s anda d ques ion is among he h ee answe s wi h highe
deg ee o ce ain y. These h ee answe s should be
p esen ed o he use . The co ec answe mus be among
hese h ee op ions. This is ela ed o p ecision, hough i
does no ma ch ha exac de ini ion ei he .
To achie e hese goals, i will be necessa y o modi y
he assigned coe icien s in p inciple. These modi ica ions
will be ca ied ou again using a bo om-up design, so ha
he ce ain y e u ned by he uzzy engine is a leas o 0.7
o he co ec decision in he lowes le el, and o 0.5 o
o he le els. Likewise, in o de o limi he numbe o
possible answe s, inco ec ques ion indexes may be
modi ied making hem lowe .
This me hod has been p o en conside ing he mos
equen ques ions asked in he po al web adminis a o as
ou se o knowledge, which consis s on a se o 117
ques ions. The elemen s o be ound a e he ques ions
hemsel es and possibly ela ed ques ions.
Tes esul s o s anda d ques ion ecogni ion i in o
i e ca ego ies:
1.-The co ec ques ion is he only one ound o he one
ha has he highes deg ee o ce ain y.
2.-The co ec ques ion is be ween he wo wi h he highes
ce ain y o is he one ha has he second highes deg ee o
ce ain y.
3.-The co ec ques ion is among he h ee wi h he highes
deg ee o ce ain y o is he one ha has he hi d highes
ce ain y.
4.-The co ec ques ion is ound bu no among he h ee
wi h he highes deg ee o ce ain y
5.-The co ec ques ion is no ound.
3.2. Rule de ini ion
As men ioned abo e, ule de ini ion co esponds o he
adminis a o . Logically, he mo e inpu s he engine has,
he mo e ules he e a e. Fo example, o a h ee inpu
engine, he inpu s can ake h ee alues: LOW, MEDIUM
and HIGH. The ou pu s can ake he alues o LOW,
MEDIUM-LOW, MEDIUM-HIGH and HIGH. The
in e ence ules de ined a e:
I all inpu s a e LOW, ou pu is LOW.
I one inpu is MEDIUM and he o he s a e LOW, ou pu
is MEDIUM-LOW.
I wo inpu s a e MEDIUM and he o he s a e LOW,
ou pu is MEDIUM-HIGH.
I all he inpu a e MEDIUM o one inpu is HIGH, ou pu
is HIGH.
The possible combina ions gene a e 27 ules o he
uzzy engine. A i e inpu engine gene a es 243 ules.
3.3. Inpu de ini ion
Fo e e y ques ion, 3 o 5 key wo ds a e de ined.
Ne e heless, he use may include in his consul a ion
anywhe e om 1 o 5 o hese wo ds. De ining an engine
wi h only a ew inpu s causes apid sa u a ion o he
sys em. This is a g ea handicap o p ecision: 90 % o
co ec answe s a e de ec ed bu only hal o hem a e he
i s op ion as may be seen in Table 1. De ining a i e
inpu engine p oduces alues wi h a e y low deg ee o
ce ain y. P ecision g ows o 55 % bu ecall alls.
A solu ion o his p oblem is he use o a iable
h esholds. When all e en ual ou pu s a e below he ixed
h eshold, his h eshold goes down un il an ou pu is
ound. I h esholds a e a iable, he sys em is mo e
lexible and esul s a e be e .
Finally, he solu ion p o ided is o implemen a lexible
sys em wi h a iable inpu s. I he use consul a ion has a
he mos h ee key wo ds, a h ee inpu uzzy engine is
used, whe eas i he use consul a ion includes ou o
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007

mo e key wo ds, he sys em will use a i e inpu uzzy
engine. Resul s a e much be e in e ms o ecall and
p ecision. The use ob ains he co ec answe 97.75 % o
he imes and he i s op ion 77.45 % o he imes.
Type o sys em /
Tes esul
ca ego y
Fi s
answe
Among
he i s
wo
answe s
Among
he i s
h ee
answe s
Ou o
he i s
h ee
answe s
Failed
answe
3 inpu engine 45 % 24 % 9 % 12 % 10 %
5 inpu sys em
wi h ixed
h esholds
55 %
12 %
3 %
1 %
29 %
5 inpu sys em
wi h a iable
h esholds
70 %
14 %
3 %
1 %
12 %
Va iable inpu
sys em wi h
ixed h esholds
77 %
16 %
4.5 %
1 %
1.5 %
Table 1. Tes esul s.
4. Conclusions
A me hod o consul a ion based on a uzzy logic
applica ion p o ided wi h an in e ace ha one may
in e ac wi h in na u al language has been p esen ed. The
sys em akes ad an age o con e ing any kind o objec in
a ex objec ; his allows he applica ion o ex e ie al
echniques. The o he ad an age o he me hod a ises om
uzzy logic lexibili y, which makes i possible o ha e a
non- igid e m weigh ing in he s age o classi ica ion o
he con en s in he se o knowledge.
A me hod o he classi ica ion o he con en s in a se
o knowledge is also p oposed. We also p esen a
modi ica ion o he classical ec o space model o de ine
e m weigh s, which a e used as inpu s o a uzzy logic
based sys em.
An e en ual applica ion o his sys em is he
implemen a ion o an in elligen agen o manage he
in o ma ion con ained in an in e ne po al. So a he
esul s ob ained a e good enough, as he numbe o
co ec ly de ec ed consul a ions is high.
5. Acknowledgmen s
The wo k desc ibed in his pape has been suppo ed by
he Spanish Minis y o Educa ion and Science (MEC:
Minis e io de Educación y Ciencia) h ough p ojec
e e ence numbe DPI2006-15467-C02-02.
6. Re e ences
[1] K.L. Kwok. "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. 1989
[2] G. Sal on. In oduc ion o Mode n In o ma ion Re ie al.
McG aw-Hill, 1983.
[3] C.J. an Rijsbe gen, . In o ma ion e ie al. Bu e wo hs,
1979.
[4] G. Sal on, C. Buckley. “Te m Weigh ing App oaches in
Au oma ic Tex Re ie al“. Technical Repo TR87-881,
Depa men o Compu e Science, Co nell Uni e si y, 1987.
In o ma ion P ocessing and Managemen Vol.32 (4), p. 431-443,
1996.
[5] G. Sal on. Au oma ic Tex P ocessing. Addison-Wesley
Publishing Company, 1988.
[6] D.L Lee, H. Chuang; K. Seamons. “Documen anking and he
ec o -space model“. So wa e, IEEE. Vol. 14, Issue 2, Ma /Ap
1997 p. 67 – 75
[7] S. Liu, M. Dong; H. Zhang, R. Li, Z. Shi. “An app oach o
mul i-hie a chy ex classi ica ion“. In e na ional Con e ences on
In o- ech and In o-ne , 2001. P oceedings. ICII 2001 - Beijing.
Vol 3, Oc / No . 2001 p. 95 – 100.
[8] M. Lu, K. Hu, Y. Wu, Y. Lu, L. Zhou. “SECTCS: owa ds
imp o ing VSM and Nai e Bayesian classi ie “. 2002 IEEE
In e na ional Con e ence on Sys ems, Man and Cybe ne ics, Vol.
5, Oc . 2002 p. 5
[9] Y. Zhao, G. Ka ypis. “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. 8-10 Ap il 2002
Page(s):16 – 21.
[10] B. Ma ín del B ío, A. Sanz Molina. Redes neu onales y
sis emas bo osos. Ra-Ma, 2001.
[11] T. Bouaziz, A. Wolski, “Applying Fuzzy E en s o
App oxima e Reasoning in Ac i e Da abases“. P oc. Six h IEEE
In e na ional Con e ence on Fuzzy Sys ems (FUZZ-IEEE'97).
July 1-5, Ba celona, Spain.
[12] M.E. Ruiz, P. S ini asan. “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.
Fou h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y (FSKD 2007)
0-7695-2874-0/07 $25.00 © 2007