Metadata
The problem in one line
Paris, paris, Paris, France, PARIS 11e, Paris (FR). To you, one place. To your catalog, five distinct values — so five search facets, five partial result sets, and one image in five missing from the count.
Normalization means reducing those variants to a single reference value while keeping a record of what was originally written.
Where the spread comes from
- Successive imports, made at different times under different conventions.
- Tools passed through: each writes the same notions into slightly different fields.
- Multiple hands: two people never type a proper noun the same way.
- Free-text entry, which permits every spelling and suggests none.
An older collection accumulates all four. That’s normal, and harmless — provided you address it before the volume makes the cleanup discouraging.
What normalization actually covers
- Case and whitespace: the easiest, and already a good share of the noise.
- Spelling variants of one entity, reduced to a preferred form.
- Synonyms, linked so that searching one finds the other.
- Equivalent fields from different tools, aligned onto a target field.
- Disguised empties:
n/a,unknown,-,tbd, which pollute facets by posing as data.
The rule that prevents damage
Normalizing doesn’t mean overwriting. The original value is information: it documents what was written, by whom, and when. A heritage collection needs that.
So keep the preferred form and the entered form. An alias pointing at a reference value fixes retrieval without destroying the record.
In practice
- Start by measuring: list the distinct values of a field, sorted by frequency. The tail of that list is your work.
- Handle case and whitespace first — immediate, risk-free gains.
- Build a vocabulary of preferred forms, using aliases rather than destructive replacements.
- Eliminate disguised empties.
- Add validation at entry, or the spread returns from the top.
Obscura Catalog attaches synonyms to a thesaurus term, so searching for one spelling also finds the others, and renaming or deleting a term does not change the keywords already set on photos — the original form stays readable after normalization.