To search for entities and properties in knowledge graphs, one can imagine lots of differents kinds of search indices to be useful. WIth GRASP (see https://grasp.cs.uni-freiburg.de), we currently use keyword search for entities and vector search for properties. However, as of now these choices were never empirically evaluated and justified compared to other types of searches.

This project is about:

• Implementing and evaluating different kinds of search for entities and properties for GRASP
  • Vector search and its variants (brute force, approximate nearest neighbor, …)
  • Traditional text search and its variants (keyword, bm25, …)
  • Hybrid search (vector + bm25, …)
• Implementing and evaluating different kinds of embeddings for entities and properties when using vector search
  • Text embeddings of labels and aliases (currently used)
  • Graph-enhanced embeddings (TransE and its variants)