Knowledge Graph

Facts written as triples — subject, relation, object — so that "Ada Lovelace" and "born in" and "London" become an edge a machine can traverse. It is the oldest idea in this corpus and the one a reader is most likely to have met outside it, because a great deal of the structured web is one.

What it buys over free text is that a query is a traversal rather than a search: the answer to a two-step question is a path, and the path is auditable. What it costs is that everything has to be extracted and typed first, and that the schema someone chose is now load-bearing — a relation your ontology does not have is a fact your graph cannot hold, however plainly the source states it.

Real graphs are extremely sparse, and this is the fact that governs everything built on them. With n entities and r relation types there are r·n² possible triples and the number asserted is a vanishing fraction of that, so absence carries almost no information. Treating an unasserted triple as false — the closed-world assumption — is therefore wrong at scale, and every method here either makes an open-world assumption or samples negatives it has no right to call negative.