Leveraging ontologies for flexible access to graph-structured data

Abstract

Knowledge Graphs (KGs) are datasets consisting of interconnected labeled entities in a sharable format useful to make different kinds of knowledge available to both humans and machines. Due to their incompleteness, a challenging task in many KG applications is that of query answering, that is, the problem of finding all possible answers to a given query. Ontologies are logical formalisms used to enrich KGs with human expertise and common-sense knowledge and offer several advantages including more complete answers to queries, by means of logical reasoning, however they cannot help with missing links between entities. In this thesis we propose an exploratory framework which leverages ontologies to support a) query formulation, allowing the user to approximate the information needs by writing a query template to describe a large set of semantically related queries, b) efficient retrieval of complete answers for the large set of related queries, c) on-the-fly query refinement, which enables interactive exploration of the data. For that we extend a well-known ontology language DL-LiteA with complex role inclusions. As a second main contribution, we present a complete complexity picture of several DL-LiteA extensions and consider the safe integration of aggregation into both the ontology and query language to support data analytics. We also propose solutions for assumption-based query answering, in which queries are equipped with assumption patterns meant for describing multiple hypothetical extensions of the KG and construct more informative answers over all such extensions. We show that assumption-based query answering is tractable in data complexity and propose ontology-based rewriting techniques for constructing conditional answers, also in the presence of closed predicates, a form of completeness statements about relations. Lastly, we also consider embedding-based ontology-mediated query answering over incomplete KGs. For that, we build on some state-of-the-art embedding models, tailored for predicting plausible answers to queries, and explore some means to incorporate ontologies, either in the training data or in the training objective function, to obtain higher accuracy for predicting missing answers that require both inductive and deductive reasoning.