Knowledge Graph Embeddings (KGE)
Traditional approaches sometimes ignore knowledge graphs’ two forms of information: high-level concepts about the general structure and detailed individual entities. These approaches usually treat all knowledge network nodes as vectors in one hidden space.
Concept2Box acquires dual geometric embeddings from a two-view knowledge graph with
(1) an ontology-view containing high-level concepts and meta-relations,
(2) an instance-view containing specific, detailed instances and relations,
(3) a collection of connections (cross-view links) between these two views. This method represents concepts as geometric boxes in latent space and things as point vectors.
Concept2Box is an alternative to a single geometric representation that cannot reflect structural differences between two views within a knowledge network and lacks probabilistic meaning in connection to concept granularity. This novel method embeds both knowledge graph views using dual geometric representations. Box embeddings teach hierarchical structures and complex interactions like overlap and disjointness.
AiThority Interview Insights : AiThority Interview with Babak Pahlavan, Founder and CEO of NinjaTech AI
These boxes’ volume matches concept granularity. In contrast, entities are vectors. A novel vector-to-box distance metric is presented to connect concept box and entity vector embeddings, which are learned together. Concept2Box was shown effective in experiments on the public DBpedia knowledge graph and a new industrial knowledge graph.
[To share your insights with us, please write to sghosh@martechseries.com]
Comments are closed.