Skip to content

Section 2 notes: ontology and corpus networks as maps

Open — needs literature research.

Key claims to support

  1. Public ontologies encode concept topology: typed identity, typed relationships, cross- ontology alignment. They are maps, not indexes.
  2. Ontological anchoring (NL goal → concept node) gives the agent a starting point, typed neighbors, and a traversal direction.
  3. Relevant path depth in well-maintained ontologies is empirically short (≤ 3–4 hops for domain-related concept pairs).

What to research

  • Wikidata and DBpedia shortest-path statistics for domain-relevant concept pairs
  • BFO adoption and coverage surveys: how many concepts in representative domains (causal inference, hydrology, economics) are represented in BFO-aligned ontologies
  • Entity linking and ontological grounding literature: how well do NL goal expressions map to ontology nodes (precision, recall, confidence)
  • SPARQL-based structured retrieval vs. embedding retrieval on KGQA benchmarks
  • OWL / SKOS / RDF ecosystem overview — what public structured knowledge is actually available and how complete is it

Candidate sources

  • Wikidata Statistics (wikistats.wmcloud.org) — concept counts, edge counts by type
  • "Knowledge Graph Question Answering" (KGQA) survey papers
  • SPARQL over Wikidata: published query complexity and result quality studies
  • BFO documentation and adoption papers (Arp, Smith et al.)
  • Entity linking papers: BLINK, GENRE, ELMo-EL, ReFinED
  • Papers on ontology-augmented RAG (a growing area — search "ontology RAG" "KG-RAG")

Key distinction to establish

Ontology ≠ knowledge base ≠ embedding index. An ontology provides relational structure (what relates to what, and how). A knowledge base populates that structure with instances. An embedding index enables similarity search over a flat representation. The paper's contribution is in the first and second categories — the relational structure is what makes directed traversal possible.

Open questions

  • How well do current LLMs perform ontological anchoring zero-shot (without explicit ontology access)? If LLMs already "know" the structure of Wikidata implicitly, the external ontology requirement is lower.
  • What is the coverage gap: what fraction of real user goals in a given domain have anchorable concepts vs. requiring open-ended synthesis?