Raw language ledger
Standardize category labels without erasing what people wrote.
Short answer: count mechanically equivalent strings together only after Unicode, whitespace, and case normalization. Keep every display form. Merge semantic synonyms manually, record the chosen standard, and make that choice reversible.
Mechanical normalization is not semantic interpretation
The field lab treats “Plan a Visit,” “plan a visit,” and extra-whitespace variants as one normalized key while preserving the observed display forms and counts. It does not automatically merge “Before you come,” “Visit planning,” and “Practical information,” because that would insert an unreviewed semantic judgment.
| Operation | Automatic? | Why |
|---|---|---|
| Unicode normalization | Yes | Provides a stable text representation. |
| Trim and whitespace folding | Yes | Avoids counts split only by accidental spaces. |
| Case folding | Yes | Avoids counts split only by capitalization. |
| Spelling correction | No | Could alter intended language or dialect. |
| Synonym merging | No | Requires contextual researcher interpretation. |
| “Better” label recommendation | No | Frequency is not proof of clarity or navigation performance. |
A defensible standardization workflow
- Export or inspect the raw-label ledger for one frozen edition and inclusion set.
- Read labels alongside the cards inside each group. A phrase can mean different things in different groupings.
- Define the purpose of standardization: reporting, comparing variants, or drafting a vocabulary.
- Enter a standard only when the equivalence is defensible for that purpose.
- Keep the raw forms, group-use count, and distinct-response count visible.
- Revisit standards when the included set changes or a second researcher disagrees.
- Export the ledger with the analysis so another person can reproduce the interpretation.
Frequency is not label quality
A common label may be vague, copied from the instructions, or driven by familiar industry language. A rare label may be especially clear. Test proposed navigation labels with realistic content and tasks rather than promoting the most frequent phrase automatically.
Do not “clean” language invisibly. Spelling, dialect, multilingual phrasing, and participant vocabulary can be part of the evidence. The app rejects unsafe control/bidirectional characters for file integrity, but it does not police ordinary language.