Knowledge graph

A knowledge graph represents subjects and the relationships among them in a form that supports connected queries. Statements may describe types, attributes and links to other entities. A graph can be expressed using RDF or other graph models; the term alone does not imply a particular storage technology, ontology or level of factual reliability. Edges are still claims whose meanings depend on definitions and source context.

In practice, graphs help answer questions that cross record boundaries, such as which organizations sponsor trials associated with a compound. Stable identifiers and clear relationship definitions are essential. The same visible link may mean ownership, collaboration or mention, each with different consequences. A graph assembled from extraction also needs a way to represent uncertainty, time and provenance alongside the connection.

For defensible use, a path through the graph is a navigational aid, not proof that every edge is valid. Each material relationship should be traceable to evidence or a declared inference. Knowledge-base population supplies statements to a graph; ontology specifies meanings; entity resolution determines when nodes denote the same real-world thing.

A graph guarantees structure, not truth. An edge between two nodes says that someone asserted a relationship; it says nothing about whether the assertion was correct, current or well sourced unless that information is stored with it. For that reason a graph used in decisions needs provenance on its statements, and the recent RDF 1.2 work on reifiers and triple terms is aimed at making such annotation a first-class feature; the article on RDF 1.2 provenance walks through what changes. The article on knowledge graph quality assessment argues that quality must be assessed against the decision the graph supports, not in the abstract.

Graphs also change. Identifiers must survive corrections, as discussed in the article on identifier versioning after corrections, and conclusions that depended on a retracted statement should be retracted with it, which is the subject of truth maintenance.

A knowledge graph is a container and a style of representation; what it contains is governed by neighbouring decisions. The ontology defines the vocabulary, entity resolution decides when two nodes are one, knowledge-base population supplies the statements, and the open-world assumption governs how a missing statement is read. A graph that stores statements without provenance can still answer questions, but cannot defend its answers, and the difference shows only when someone asks for the basis of a particular edge.