TensorKG is an experimental GPU-oriented reasoning and scored concept alignment engine for OWL-backed KGs. It compiles OWL class restrictions into a directed acyclic graph of constraints, converts KGs into a GPU-friendly tensor representation, and then evaluates each layer of constraints against all KG nodes simultaneously on the CPU or GPU (via CUDA bindings).
TensorKG is primarily useful for two tasks: materialization of type assignments at scale, and alignmemt of KG nodes to ontology classes at scale—essentially, scoring how well each node fits the structural requirements of each class. At present, TensorKG fully supports OWL EL semantics, and partial support is provided for some of OWL DL semantics as well. For more information, see the GitHub repository: https://anonymous.4open.science/r/tensorkg-anonymized/README.md .
- Forward reasoning: Materialization of type assignments from sufficient conditions. In other words: it answers the question "Can we infer node n must be an instance of concept C from known information?" for all nodes in the graph and concepts in the ontology.
- Admissibility testing: Checking which KG nodes satisfy the necessary conditions of which classes, according to known information. In other words, it answers the question "Would asserting node n as class C be consistent with everything we know about n and C?" for all nodes in the graph and concepts in the ontology.
- Scored semantic alignment: The score-based version of admissibility testing: scoring the degree to which each KG node satisfies the necessary conditions of each class. In other words, for some node n and class C, a score s(n,C)=0.0 means n is not known to satisfy any conditions of C, and a score s(n,C)=1.0 means n is known to satisfy all known requirements of C.