Probabilistic Reasoning in the Semantic Web using Markov Logic
Authors
Abstract
The Semantic Web envisions a world where agents share and transfer structured knowledge in an open and semi-automatic way. In most of the cases, this knowledge is characterized by uncertainty. However, Semantic Web languages do not provide any means of dealing with this uncertainty; they are mainly based on crisp logics, unable of dealing with partial and incomplete knowledge. Reasoning in the Semantic Web resigns to a deterministic process of verifying if statements are true or false.In the last years, some efforts have been made in representing and reasoning with uncertainty in the Semantic Web. These works are mainly focused on how to extend the logics behind Semantic Web languages to the probabilistic/possibilistic/fuzzy logics, or on how to combine these languages with probabilistic formalisms like Bayesian Networks. In all of these approaches, this is achieved by annotating the ontologies with some kind of uncertainty information about its axioms, using this information to perform uncertainty reasoning. Nevertheless, several questions arise: how can reasoning be done efficiently with this uncertainty information? Where to get this uncertainty information?
In this thesis, we present solutions for both questions. The solution for the first question is Markov logic, a new promising approach to reasoning with uncertainty. In this type of logic, there is no right and wrong world; there are multiple worlds with different degrees of probability. This is done by combining logic and probability in the same representation, and then using efficient learning and inference algorithms. For the second question, several solutions were developed:
· If the ontologies are annotated with some kind of uncertainty information, like probabilities, Markov logic can be used to reasoning about this information;
· If the ontology contains individuals, those individuals can be used to automatically learn the uncertainty of the ontology;
· If the ontology does not comprise uncertainty information or individuals, both resources can be automatically learned by analyzing textual resources and web search engines.
We developed a system, Incerto, which explores the capabilities of Markov logic for the Semantic Web. This system was applied in several interesting tasks, like reasoning about automatically learned ontologies and social networks analysis. The main contributions of this thesis are:
· The application of Markov logic for learning and reasoning about uncertainty in the Semantic Web;
· The development of several techniques for learning automatically the uncertainty of ontologies;
· The development of a new technique to parameterize Markov logic networks with probabilities;
· The development of a new technique to learn the probability of ontology axioms by using web search engines;
· The development of Incerto, and its application to several Semantic Web domains.