Research

Ontology and Knowledge Representation

Applied Ontology is an area of symbolic artificial intelligence that aims to develop high-quality,  reusable, and maintainable representations of a domain. For this purpose, it utilizes philosophical methods and theories in practical, real-world contexts. Our research covers a wide range of topics in the area of applied ontology, including the philosophical foundations of ontologies, their relationship to cognitition (e.g., image schema),  and methodologies for ontology development and evaluation (e.g., semantic dependency). Further, we develop ontologies in cooperation with domain experts.

 

Semantic methods & formal ontologies

The main insight is that ontologies, models, and specifications can be seen as heterogeneous modular logical theories in some logical system. For complex systems, an early formal model can be checked early for consistency, resulting in possible design changes – instead of doing this later (which means more involved and costly) in the development. We have successfully applied this to ontologies and specifications.

The specification and modeling language DOL ("Distributed Ontology, Model and Specification Language") has been developed within the OMG, led by Till Mossakowski and Fabian Neuhaus. We develop a suitable language with formal semantics, as well as proof methods and tools, among them the Heterogeneous Tool Set.

Neuro-symbolic integration

Analogous to human information processing that integrates more rational, slow reasoning and more holistic and intuitive, faster information processing, neural-symbolic learning approaches integrate logical reasoning with neural networks. In this way, expert background knowledge (in our group, mainly ontologies and their logical axioms) can be used by neural networks in order learn both from data and axioms, thus improving performance while simultaneously reducing the amount of data that is needed for training.

While in many neural-symbolic approaches, logical and neural modules cooperate, we focus on tightly integrated approaches like conceptors and logical neural networks, because these are most promising for a new generation of artificial intelligence systems. More...