A Method for Modeling Uncertainty in Semantic Web
Taxonomies
Markus Holi and Eero Hyvönen
University of Helsinki,Helsinki Institute for Information T echnology (HIIT)
P .O.Box 26,00014UNIVERSITY OF HELSINKI,FINLAND
引号的使用
www.cs.helsinki.fi/group/co/,firstname.lastname@cs.helsinki.fi
World 37*23 = 851Europe 15*23 = 345Asia 18*23 = 414EU 8*21 = 168Sweden 4*9 = 36Finland 4*9 = 36Sweden
Finland
EU
Europe Asia
World
Lapland
Norway水浒传又名什么
Norway 4*9 = 36Lapland 13*2 = 26Russia
Russia 18*19 = 342
新的旅程
Russia&Europe = 57Russia&Asia = 285
Lapland&(Finland | Sweden | Norway) = 8Lapland&Russia = 2 Lapland&EU = 16
Figure 1:A Venn diagram illustrating countries,areas,their
overlap,and size in the world.
ABSTRACT
We prent a method for reprenting and reasoning with uncer-tainty in RDF(S)and OWL ontologies bad on Bayesian networks.94年属
1.UNCERTAINTY IN ONTOLOGIES
Taxonomical concept hierarchies constitute an important part of the RDF(S)1and OWL 2ontologies ud on the mantic web.For example,subsumption hierarchies bad on the subClassOf or partOf properties are widely ud.In the real world,concepts are not always subsumed by each other,and cannot always be orga-nized in crisp subsumption hierarchies.Many concepts only partly overlap each other.See,for example,the Venn diagram of fig-ure 1illustrating various countries and areas in the world.A crisp partOf meronymy cannot express the simple fact that Lapland par-tially overlaps Finland,Sweden,Norway,and Russia,nor quantify the overlap and the coverage of the areas involved.
Semantic web ontologies are bad on crisp logic and do not usually provide well-defined means for expressing degrees of sub-sumption.To address this foundational problem,this paper prents a new probabilistic method to model conceptual overlap in tax-onomies,and an algorithm to compute the overlap between a -lected concept and the other concepts of a taxonomy.Our approach
Figure3:The taxonomy offigure2transformed into the solid path structure(Bayesian network).The original partial inclu-
sions of Lapland and Russia is transformed into crisp subsump-刘英雄
tion by using middle concepts.Note that disjoint concepts are
鸡汤文
d-parated.
3PUTING OVERLAPS
Given a taxonomy we want to know how much the concepts have
in ,overlap with each other.For example,assume that
one is interested in a concept.We want a method to evaluate,how
much the other concepts in the taxonomy have in common with .is called the lected concept and the evaluated concepts are
called referred concepts.Let be a referred concept.The question of how much has in common with c
万古江河
an be quantified in well-defined n in terms of the conditional probability
.This probability is bad on the t theoretic structure of the
taxonomy.
In theory,the conditional probability can be computed directly
from the Venn diagram.In practice,this is complicated,inefficient,
and the Venn diagram may not be available.To solve the problem,
we have developed an algorithm for transforming the RDF(S)graph
into a Bayesian network.After this,the efficient evidence propa-
蜡油
gation algorithms developed for Bayesian networks can be ud for computing the needed probabilities.We briefly describe next how this can be done.
The overlap value between concepts(lected)and(re-ferred)is
.If,then.
As can be en,the topology of the solid path structure is well-suited to be ud as a Bayesian network.Let(lected)and (referred)be concepts with the overlap value.Probabilistically and are boolean random variables,and.
The conditional probability table(CPT)for each node can be constructed in the following way:1)Go through all the value com-binations of the parents of.2)The true value in the CPT for a given entry is
3www.swi-prolog/
/
