Why a place called Lakeview Park is not a lake

If you are building a dataset of natural places and you have 53,082 of them, the tempting shortcut is to classify by name. It is fast, it needs no external data, and it is wrong often enough to be dangerous.

The failure mode

Name-based classification fails in both directions, but the false positives are worse. “Lakeview Park” contains the word lake. So does “Lakeside Industrial Estate”. Match on the substring and a business park joins your lakes dataset, where it will sit quietly until a reader notices — and by then the reader has stopped trusting everything else on the page.

The reverse case is more common but less damaging: a lake named after a person, with nothing in its name to indicate what it is, simply goes missing. An absent lake is a gap. A car park listed as a lake is a credibility problem.

What we do instead

Classification comes from the structured type. Each place carries a Wikidata P31 statement — “instance of” — and we follow at most one P279 “subclass of” hop to catch the cases where something is typed as a specific kind of a broader category.

The one-hop limit is deliberate. Follow the subclass chain far enough and almost everything eventually becomes a geographic feature, at which point the taxonomy has stopped distinguishing anything.

Where names are still useful

Names get exactly one job, and it runs in one direction only: a name may remove a category, never add one.

A structured type is a claim about what something is, and it is usually right. But it is sometimes right in a way that is misleading in context — a feature technically typed as a body of water that is, in practice, a drainage basin nobody would visit. The name is often the fastest signal that a technically-correct type is contextually wrong. So a name can veto a classification the type would otherwise allow.

It can never do the opposite. Nothing named “Lakeview” becomes a lake here, no matter how convenient that would be for the size of the lakes dataset.

The general principle

The direction of the rule matters more than the rule. Evidence promotes; heuristics may only demote. That ordering means the worst case for a heuristic error is a missing entry — which is recoverable — rather than a fabricated one, which quietly costs you the reader.