Each entity gets a learned-style embedding vector. A neural module scores every unlinked pair by cosine similarity; pairs that clear the confidence threshold become proposed links. A symbolic module then reasons over the trusted graph (curated facts + accepted neural links) with a transitive rule, chaining edges into new derived facts.
sim(u,v) = cos(e_u, e_v)
c(u,v) = (sim(u,v) + 1) / 2
accept neural edge if c(u,v) ≥ τ
R(a,b) ∧ R(b,c) → R(a,c) (chained up to d hops)
- Entities — nodes in the knowledge graph, coloured by type.
- Confidence τ — how strict the neural module is before it trusts a similarity-based link.
- Reasoning hops — how many transitive steps the symbolic engine is allowed to chain per query.
- Neural / Symbolic / Hybrid — isolate each reasoning layer or watch them combine.
This is the core idea behind neuro-symbolic AI: neural nets are good at fuzzy pattern-matching (embeddings, similarity), while symbolic logic is good at exact, explainable multi-step inference — combining both lets a system both "notice" a likely connection and "prove" a derived one, e.g. for LLM + knowledge-graph reasoning pipelines.