Learn · Beginner

Knowledge graphs: storing facts as named things and explicit relationships

A knowledge graph stores information as identifiable things connected by explicit relationships. It lets a system ask how people, organizations, documents and events relate, instead of relying only on matching words or finding similar passages. Its value comes from making connections inspectable; storing an assertion in a graph does not make that assertion true.

From sentences to relationships

Consider three statements: a researcher wrote a paper, a laboratory announced it, and another team checked a theorem from it. A paragraph can express all three, but a computer needs to distinguish the people, organizations, document and verification event before it can reliably answer who did what.

A graph represents those things as nodes and their relationships as edges. A simple fact can be written as a triple: subject, predicate, object. For example, “Paper A — authored by — Researcher B.” A second triple might say “Paper A — announced by — Laboratory C.” The relationship names preserve distinctions that a vague list of associated names would lose.

Think of a carefully labeled map. Streets connect particular places in particular ways. Two places being near each other on the page does not mean a road joins them. Likewise, two organizations appearing in the same article does not establish a collaboration, an ownership relationship or an endorsement.

Names and definitions do real work

Entity identity is the first challenge. The same person may appear under initials and a full name; two different people may share a name. A reliable graph assigns stable identities and keeps evidence for deciding which mentions refer to the same thing. Incorrectly merging two people can contaminate every query that follows their relationships.

The relationship vocabulary matters just as much. “Funded,” “authored,” “hosted” and “reviewed” mean different things. A schema defines the kinds of entities and relationships the system accepts. It can prevent obvious category mistakes, such as treating a publication date as a person, but cannot prove that a correctly formatted statement is accurate.

Maximilian Nickel and colleagues describe symbolic relations and learned relational models in A Review of Relational Machine Learning for Knowledge Graphs. Aidan Hogan and colleagues’ Knowledge Graphs surveys the broader foundations, including representation, construction and use. Together they show why a graph is both a data structure and a collection of modeling decisions.

Following a path to an answer

Suppose a newsroom wants every paper announced by a laboratory and independently checked by another institution. A graph query can follow an announcement relationship to a paper, then a verification relationship to a checking event and its institution. The answer comes from a specified pattern of connections.

That structure also exposes missing evidence. A paper might have an author’s own check but no independent verification relationship. The query should not quietly treat those as equivalent. Nor should absence of a relationship automatically mean the event never happened: the graph may simply be incomplete.

Today’s AI-assisted Ghys result illustrates the distinction. Authorship, model assistance and formalization are separate claims with separate participants. Representing them explicitly helps prevent a system from turning “a laboratory announced a model-assisted paper” into “the laboratory’s model autonomously proved the entire conjecture.” The graph supports careful retrieval; the underlying paper supplies the evidence.

Graphs and language models

An embedding places a passage or entity in numerical space so similar items can be retrieved. A knowledge graph explicitly records how named entities relate. Similarity is useful when a user describes something vaguely; explicit relations are useful when the question depends on a precise connection.

A retrieval-augmented system can combine them. Similarity search finds relevant documents or candidate entities; graph queries gather connected facts; the language model explains the result with references. That architecture still needs checks that retrieved evidence supports the answer. Fluent prose cannot repair an incorrect relationship.

Learned graph embeddings can also score missing relationships. Such predictions are hypotheses, not newly verified facts. A model may infer a likely affiliation from patterns in existing data, but publishing it as confirmed would confuse statistical plausibility with evidence. Keep predicted edges distinguishable from sourced assertions.

Evidence, time and limitations

Useful graphs record where a claim came from and when it applied. A person can change employers, a model can gain a new version, and a proposed fund can await approval. A timeless ownership edge could misrepresent a conditional announcement. Conflicting sources may require multiple attributed assertions rather than silently overwriting one with another.

Graphs also cost effort to maintain. Extraction tools can misread text, schemas can impose misleading categories, and missing relationships can bias the answers. Graph neural networks are a different layer: they learn from graph structure, whereas a knowledge graph organizes asserted information whether or not a neural model is used.

Use a knowledge graph when important questions depend on explicit relationships, stable identities and traceable evidence. Use documents alongside it when qualifications and context matter. The practical discipline is simple: preserve who asserted what, about which entity, under which conditions, before asking a model to turn those connections into an answer.

Ground Truth. (2026, October 19). Knowledge graphs: storing facts as named things and explicit relationships. Ground Truth. https://groundtruth.day/learn/knowledge-graphs.html

@misc{groundtruth:knowledge-graphs,
  title  = {Knowledge graphs: storing facts as named things and explicit relationships},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/learn/knowledge-graphs.html}
}

Topics: fundamentals · knowledge-graphs · retrieval · data · reasoning