Implementing TripleStoreNeuralReasoner (Embedding Based Reasoner)

[Demirrr]

TripleStoreNeuralReasoner should be a part of the owlapy.

class EBR(AbstractOWLReasoner)

• Suggest the init arguments that requires an instance of NeuralOntolgy @LckyLke
• Implement the Interface, abstract methods
• Implement named concept and negation
• Implement union and intersection
• Implement existential
• implement universal
• Implement cardinality restrictions
• Discuss about integration data properties

[LckyLke]

https://github.com/dice-group/owlapy/blob/45f1ad6be77d09e14c18f0b69ee17741163f8eef/owlapy/abstracts/abstract_owl_reasoner.py#L20C1-L27C1

The abstractOWLReasoner interface requires giving an ontology, which we do not need for the embedding based approach.
→ we just need an embedding path
ig we could define a new "neuralOntology" but idk how sensible that is

[Demirrr]

Yes, abstractOWLReasoner indeed requires an instance of AbstractOWLOntology.
We can implement a new class called neuralOntology that inherits from AbstractOWLOntology.

neuralOntology requires a path of a pretrained KGE embedding model and implement the abstract classes.

[LckyLke]

do you guys think it would make sense to batch certain calls:
Consider a call ∃x.hasGender(x,Female) -> for this we would first retrieve all females (ie the filler in that ce) and then would have to make the same prediction for each filler object (here with r being has Gender) to get it.

Currently, we use a for loop, but i was thinking of putting all that in a tensor such that we could potentially do this on the GPU in parallel.

[Demirrr]

Please add your such suggestions on vectorizing the computation in the respective methods.
We will focus on the improvements in the refactoring phase of EBR :)

[LckyLke]

https://github.com/dice-group/owlapy/tree/EBR

I created this branch with (slightly refactored) code from ontolearn

[Louis-Mozart]

Ok, nice, so how do we share the tasks?

We should also make sure that EBR always returns proper individuals. See below an example of returned instances.

{ 'http://www.w3.org/2002/07/owl#Ontology', 'http://www.w3.org/2002/07/owl#NamedIndividual', 'http://example.com/father#markus', 'http://example.com/father#michelle', 'http://www.w3.org/2002/07/owl#Class', .....}

[Demirrr]

@Louis-Mozart you can take a look at the implementation of @LckyLke . He completed the tasks already. Perhaps you can review his changes and add comments if you have some

[LckyLke]

Yes right now I am working on the vectorization in EBR as well, so the code might still change
-> but I will finish it tmmrw probably

[LckyLke]

This also means that the current implementation only works with the batch-scoring branch of dicee

https://github.com/dice-group/dice-embeddings/tree/batch-scoring

[Louis-Mozart]

Alright, I will do that.