Taxonomy extraction using knowledge graph embeddings and hierarchical clustering

Amal Zouaq

Abstract:

While high-quality taxonomies are essential to the Semantic Web,
building them for large knowledge graphs is an expensive process.
Likewise, creating taxonomies that accurately reflect the content
of dynamic knowledge graphs is another challenge. In this paper,
we propose a method to automatically extract a taxonomy from
knowledge graph embeddings, and evaluate it on DBpedia. Our
approach produces a taxonomy by leveraging the type information
contained in the graph and the tree-like structure of an unsupervised hierarchical clustering performed over entity embeddings.
We then extend our method with an axiom induction mechanism
which allows us to identify new classes from the data and describe
them with logical axioms, thus leading to expressive taxonomy
extraction.