Friday 6 December, 2013 at 11h00 in Celestijnenlaan 200A (room 05.001)
Efficient Maximum A-Posteriori Inference in Statistical Relational Models and Applications in Description Logics
by Jan Noessner (University of Mannheim)
The maximum a-posteriori (MAP) query in statistical relational models computes the most probable world given evidence and further knowledge about the domain. It is arguably one of the most important types of computational problems, since it is also used as a subroutine in weight learning algorithms. We focus on Markov logic (ML) as statistical relational formalism. Markov logic combines Markov networks with first-order logic by attaching weights to first-order formulas.
In this talk, we discuss an improved inference algorithm and some applications for MAP queries. Our inference algorithm leverages symmetries in ML networks and parallelizes MAP inference by translating the ML network to an integer linear program (ILP). Additionally, we integrate cutting plane inference (developed by Sebastian Riedel) which significantly reduces the number of groundings by solving multiple smaller ILPs instead of one large ILP.
Afterwards, we apply the MAP query to description logics. Description logics (DL) are knowledge representation formalisms whose expressivity is higher than propositional logic but lower than first-order logic. The most popular DLs have been standardized in the ontology language OWL. We combine Markov logic, which essentially is a log-linear model, with description logics to log-linear description logics. The MAP query then allows us to compute the most-probable 'coherent' world. Possible applications of log-linear description logics are mainly located in the area of ontology learning, ontology debugging, and ontology matching. For the latter one we briefly scratch two related approaches in terminological and instance matching.