How AI Can Connect Events Across Time

Two events published four years apart, by different outlets, in different languages, describing different-looking situations, can be the same story. Recognising that is the part of research that takes the longest and that almost no tool helps with.

Why keyword search fails at this

Search matches strings. Connection is about meaning, and meaning survives none of the things that change between two accounts of related events:

• vocabulary drifts — the same entity is a “start-up”, then a “company”, then “the operator”;
• language changes — the earlier account exists only in the language it was published in;
• framing changes — an event described as a technical decision in 2022 is described as a policy dispute in 2025;
• the subject is not named — the connecting element is a person or a location mentioned in passing.

Any of these is enough to break a keyword match while leaving the relationship perfectly real.

What connects two events

There are only a handful of ways two events relate, and each is something you can extract and compare:

Shared entities. The same person, organisation, place or product appears in both. This is the most common bridge and the easiest to get wrong, because names are ambiguous — the entity has to be resolved, not just matched.

Shared topic. Both belong to the same domain, at a useful level of specificity. “Technology” connects nothing; a narrower classification connects a great deal.

Shared source lineage. Both trace back to the same original report. Useful in the opposite direction too: it tells you when apparent corroboration is really a single source repeated.

Temporal proximity with a causal shape. Events clustered in time around the same entities often stand in sequence — announcement, reaction, consequence. Proximity alone is weak; proximity plus shared entities is strong.

Semantic similarity. The two descriptions mean approximately the same thing even with no vocabulary in common. This is what language models are genuinely good at, and it is the bridge that survives translation and rewording.

The pipeline

In practice, connecting events across time is a sequence of ordinary steps:

1. Extract — pull events and entities out of unstructured text, with dates.
2. Resolve — decide that two mentions refer to the same entity, and that two accounts describe the same event.
3. Relate — compute the links above, each with an indication of how it was derived.
4. Place — position the event chronologically, using the date it happened rather than the date it was published.
5. Expose — surface the connection with its basis visible, so a reader can judge it.

Step five is not decoration. A connection you cannot inspect is a claim you have to take on faith, and this is precisely the kind of system where taking things on faith is the failure mode.

What it gets wrong

Entity resolution fails on common names and organisations that change identity. Event deduplication merges things it should not, or splits things it should not. Semantic similarity finds pairs that are topically close and causally unrelated.

None of this is a reason to skip the work — a researcher doing it by hand makes the same errors, more slowly. It is a reason to present connections as candidates with visible reasoning rather than as facts.

The result

When it works, a subject stops being a pile of documents and becomes a graph you can walk. You enter at any event and move outward: to its sources, to the entities involved, to what preceded it, to what it led to.

That walk is what research actually consists of. The point of connecting events is to let it happen in minutes rather than weeks.