Knowledge and learning: synergies between ontologies and machine learning

Martin Glauer

Abstract: The growing pool of available information and knowledge makes suitable knowledge representation indispensable. Over the past decades, ontologies have emerged as a suitable method for describing knowledge flexibly and with rich semantics. However, ontology development is a labor-intensive and time-consuming process that requires consensus-building processes involving many experts. This growth in available information is further enhanced by machine learning meth- ods. These have delivered astonishing results, particularly in recent years. However, many of them suffer from a transparency problem. In this thesis, I show synergies between the world of ontologies and that of machine learning that can address their respective problems. We will present a novel, machine- learning-based approach to extend ontologies. Existing approaches rely largely on existing domain literature, which does not reflect the consensus that inheres in the ontology. Contrary to that, our approach is based solely on an existing ontology and its annotations. I will evaluate the proposed prediction system on a case study on the ChEBI ontology and demonstrate an interface that can be used by ChEBI developers in order to accelerate the development process. Conversely, we will also present two methods with which ontologies can be used in machine learning. Firstly, I will show that an ontology can serve as a knowledge base for learning tasks, even if the knowledge relevant to the task is not explicitly represented. Furthermore, I will present a method to use the semantic structure of the ontology to improve the logical consistency of predictions.