Core Concepts
Neuro-symbolic reasoner Sandra combines vectorial representations with deductive reasoning based on the Description and Situation (DnS) ontology design pattern.
Abstract
Abstract:
Sandra combines vectorial representations with deductive reasoning.
Geometric nature allows combination with neural networks.
Introduction:
Reasoning on perspectives is relevant in various domains.
Sandra infers plausible descriptions for a given situation.
Descriptions & Situations:
DnS formalizes frame semantics, generalizing Fillmore's proposals.
DnS defines vocabulary for n-ary relations, introducing concepts of description and situation.
Method:
Defines a vector space V over R with dim(V) = |R ∪ D|.
Function fs maps a situation s ∈ S into a vector in V.
Experiments:
Tested on I-RAVEN benchmark and RotatedFashionMNIST dataset.
Integration of sandra improves interpretability without loss of performance.
Stats
Sandra builds a vector space constrained by an ontology and performs reasoning over it.
The method formalizes the inference of satisfies relation between situations and descriptions.
Quotes
"Reasoning on perspectives is relevant in many contexts."
"Sandra infers all perspectives that are plausible descriptions for a given situation."