Type vector representations from text: An empirical analysis

Knowledge Graphs (KGs) are abstractions used to represent knowledge in which real-world entities are organized using a type system where types are organized using a sub-type relation: the ontology. A key factor in many applications is to evaluate the similarity between the types of the ontology. Classical measures to evaluate the semantic similarity between types are often based on the structured organization of the sub-type system. In this work, we show that it is possible to use methods coming from Natural Language Processing to embed types in a vector space starting from textual documents. We show that in this representation some of the properties of the hierarchy are still present and that the similarity in this space captures also characteristics that are close to human behavior.