TBox learning from incomplete data by inference in BelNet⁺

In this work we deal with the problem of TBox learning from incomplete semantic web data. TBox, or conceptual schema, is the backbone of a Description Logic (DL) ontology, but is always difficult to obtain. Existing approaches either fail in getting correct results under incompleteness or learn results that are not enough to resolve the incompleteness. We propose to transform TBox learning in DL into inference in the extension of Bayesian Description Logic Network (abbreviated as BelNet⁺), whereby the structure in the data is leveraged when evaluating the relationships between two concepts. BelNet⁺, integrating the probabilistic inference capability of Bayesian Networks with the logical formalism of DL ontologies - Description Logics, supports promising inference. In this paper, we firstly explain the details of BelNet⁺ and introduce a TBox learning approach based on BelNet⁺. In order to overcome the drawbacks of current evaluation metrics, we then propose a novel evaluation framework conforming to the Open World Assumption (OWA) generally made in the semantic web. Finally the results from empirical studies on comparisons with the state-of-the-art TBox learners verify the effectiveness of our approach.