BayesOWL: A Probabilistic Framework for Semantic WebTweet
Authors: Zhongli Ding
Date: December 05, 2005
To address the difficult but important problem of modeling uncertainty in semantic web, this research takes a probabilistic approach and develops a theoretical framework, named BayesOWL, that incorporates the Bayesian network (BN), a widely used graphic model for probabilistic interdependency, into the web ontology language OWL. This framework consists of three key components: 1) a representation of probabilistic constraints as OWL statements; 2) a set of structural translation rules and procedures that converts an OWL taxonomy ontology into a BN directed acyclic graph (DAG); and 3) a method SD-IPFP based on 'iterative proportional fitting procedure' (IPFP) that incorporates available probability constraints into the conditional probability tables (CPTs) of the translated BN. The translated BN, which preserves the semantics of the original ontology and is consistent with all the given probability constraints, can support ontology reasoning, both within and cross ontologies, as Bayesian inferences, with more accurate and more plausible results.
SD-IPFP was further developed into D-IPFP, a general approach for modifying BNs with probabilistic constraints that goes beyond BayesOWL. To empirically validate this theoretical work, both BayesOWL and variations of IPFP have been implemented and tested with example ontologies and probabilistic constraints. The tests confirmed theoretical analysis.
The major advantages of BayesOWL over existing methods are: 1) it is non-intrusive and flexible, neither OWL nor ontologies defined in OWL need to be modified and one can translate either the entire ontology or part of it into BN depending on the needs; 2) it separates the 'rdfs:subClassOf' relations (or the subsumption hierarchy) from other logical relations by using L-Nodes, which makes CPTs of the translated BN smaller and easier to construct in a systematic and disciplined way, especially in a domain with rich logical relations; 3) it does not require availability of complete conditional probability distributions, pieces of probability information can be incorporated into the translated BN in a consistent fashion. One thing to emphasize is that BayesOWL can be easily extended to handle other ontology representation formalisms (syntax is not important, semantic matters), if not using OWL.
Organization: Computer Science and Electrical Engineering
School: University of Maryland, Baltimore County
Google Scholar: Ut7OZwAEwYkJ
Number of Google Scholar citations: 10 [show citations]
Number of downloads: 6760
Available for download as