Knowledge graph

A knowledge graph stores entities and the typed relationships between them, in a form built for traversal rather than retrieval. The unit is the statement — subject, predicate, object — and the value comes from the fact that statements compose: if a paper reports a result, the result concerns a variable, and the variable was measured by an instrument, then a query can reach from instrument to conclusion without anyone having anticipated that particular path.

That composability is the argument for the structure. A conventional database answers questions its schema was designed for. A graph answers questions no one specified in advance, provided the relationships are typed consistently.

What it offers research

The literature is a large set of documents with relationships that exist only in prose. Which studies used the same method, which contradict each other, which used the dataset now known to have a calibration problem — all recoverable by a person reading enough papers, none of it queryable. A knowledge graph is an attempt to move those relations into a form a machine can traverse, which is what ORKG targets for scholarly claims and what MaRESS and machine-actionable knowledge for Earth science address for this domain.

In Earth system science specifically, the appeal is that the connections between a dataset, the model that produced it, the instrument that constrained it, and the paper that interpreted it are exactly the ones people spend real effort reconstructing manually.

Why they are harder than they look

Someone has to build the statements. Automatic extraction from text is unreliable for anything requiring judgment, and manual curation does not scale to the literature. Nearly every knowledge graph project runs into this, and the honest ones are explicit that coverage is the binding constraint rather than the technology.

The ontology is a commitment. Deciding what entity types and relationship types exist means fixing, in advance, which distinctions matter — and any fixed vocabulary excludes distinctions someone will later need. This is the same problem Vervaeke’s relevance realization names on the cognitive side: relevance cannot be settled ahead of time, and a schema settles it anyway. The tension is structural, not a defect of any particular ontology.

Graphs represent assertions, not truths. A statement in a graph says a source claimed something. Without provenance attached, contested claims and settled ones are indistinguishable, and a graph that flattens the difference is actively misleading. This is why provenance is not an optional enrichment here — it is what makes the assertions interpretable.

See also: ORKG, metadata for the descriptive layer beneath it, provenance, and the open science and knowledge graphs map.