How AI Finds Relationships in Information

A knowledge graph is a simple idea with a difficult middle. Things become nodes, relationships become edges, and the result can be walked instead of read. The difficulty is not the model — it is everything required to build it from text written by people who were not thinking about graphs.

Step one: find the things

Entity extraction pulls out the people, organisations, locations, products and other named things a text refers to. Modern language models do this well enough that it is rarely the bottleneck.

What is harder is knowing what kind of thing something is. A name can be a person, a company, a place or a product depending on context, and getting the type wrong propagates into every relationship built on top of it.

Step two: decide what is the same

This is where most of the difficulty lives, and it has an unglamorous name: entity resolution.

Three mentions across three documents may refer to one thing or three:

• an organisation appearing under its full legal name, its trading name and an abbreviation;
• two people with the same common name;
• a company that renamed after a merger, whose earlier and later identities are the same continuous entity for some purposes and not for others.

Resolve too aggressively and you merge distinct things, producing relationships that do not exist. Resolve too conservatively and a long-running subject fragments into disconnected pieces exactly where continuity mattered.

There is no setting that is right in general. What helps is context: two mentions co-occurring with the same other entities, in the same domain, in a compatible time range, are more likely to be one thing than two mentions matching only on string similarity.

Step three: find the edges

With entities resolved, relationships come from several signals:

Co-occurrence. Appearing together, repeatedly, in the same events. Weak individually, meaningful in aggregate.

Syntactic relations. The sentence states the relationship directly: who acquired whom, who was appointed to what. Precise where present.

Semantic similarity. The descriptions mean similar things without sharing vocabulary. This is what survives paraphrase and translation, and what keyword systems cannot do at all.

Temporal correlation. Entities whose activity rises and falls together over time. Suggestive, and easy to over-read.

Source lineage. Two accounts deriving from the same original. Important in the negative: it tells you when apparent independent confirmation is one source repeated.

Step four: keep the reasoning attached

An edge without a basis is an assertion. The system should be able to answer “why is this here” with the specific evidence: these documents, this sentence, this pattern of co-occurrence.

This is not a nicety. Automatically derived relationships are wrong often enough that an uninspectable graph is actively misleading, while an inspectable one is useful even when a given edge is wrong — because the reader can see it is wrong.

What it is for

A graph is not the deliverable. It is the substrate that makes the rest work: it is what lets a timeline show related events, what lets a source view show which other sources touch the same entities, and what turns a subject into something you can walk rather than scroll.

The reader never has to see the graph. They just notice that following a thread takes one click instead of forty minutes.