For the second year straight, the Discovery Lab will be at ICLR! This time members of the lab will present StarQE: Query Answering on hyper-relational KG’s.

StarQE: Query Embedding on Hyper-Relational KG’s is the product of an international collaboration between Dimitrios Alivanistos (VU Amsterdam, DiscoveryLabs), Max Berrendorf (LMU Munich), Michael Galkin (Mila Quebec & McGill University) and Michael Cochez (VU Amsterdam, DiscoveryLabs).

There are many papers in the literature that experiment with graph embeddings for answering queries over knowledge graphs. In their paper, members of the Discovery Lab, together with colleagues from Germany and Canada, have extended those algorithms to a richer family of graphs, called “hyper-relational graphs”. These graphs allow for edges to be labelled with key-value pairs, instead of just a single label. This richer representation is gaining popularity as the go-to choice for public KG’s (WikiData) and in daily use in some of the knowledge graphs constructed and maintained by Elsevier.

This is the first paper to explore query answering with the use of embeddings for such hyper-relational graphs. The authors also introduced a new hyper-relational query dataset - WD50K-QE which they used for their experiments. They further analyse the impact of qualifier information in the performance of query answering systems and show a great boost when qualifiers are available!