Evaluating Different Methods for Semantic Reasoning Over Ontologies
Abstract
Reasoning over knowledge bases such as Semantic Web ontologies enables the discovery of new facts from existing knowledge. Knowledge-enhanced machine learning has motivated the development of neuro-symbolic reasoners, which enable faster but approximate computation of new facts or entailments. Neuro-symbolic methods generate vector representations (embeddings) of entities in a knowledge base. We analyze some ontology embedding methods, by implementing them as neuro-symbolic reasoners and evaluating their predictive performance on the datasets and tasks provided by the Semantic Reasoning Evaluation Challenge 2023. We explore two types of embedding methods: graph-based and modeltheoretic. Regarding graph-based embeddings, we evaluated the impact of different combinations of graph representation of ontologies with knowledge graph embedding methods. For model-theoretic embeddings, which create models for theories, we evaluate the impact of using several models, enabling approximate semantic entailment.