Unbounded tagging destroys retrieval value

What it is

Once coining is allowed, the vocabulary grows: the same meaning appears as “budget,” “Budget,” “Q3-budget,” “budget-draft.” More words make filtering less like classification and more like searching spelling. Guided convergence means, at coin and filter time, suggesting merges, frequent words, and existing near-matches so the list stays scannable—not a hard cap that blocks new words. It happens after people can already coin. Clear AND/OR will not save retrieval against two hundred near-synonyms.

Why it happens

Coining is almost free locally and expensive later for future-self and colleagues. People forget which spelling they used three months ago; ticking one hides the other half. Suggesting existing words, listing near-matches while typing, and offering “merge with X” pull that global cost into the second of coining. A hard cap (20 tags max) breeds stranger compounds, or abandonment. Convergence must reverse: a bad merge must split. Auto-merging short words with different senses (“design”) destroys retrieval, so merges need a nod. Rare words can sink in the list without being deleted from objects.

Studying it

In a compressed multi-week rhythm, have people tag similar items repeatedly. Watch vocabulary length, synonym clusters, misses when retrieving by one word. Compare no suggestions, near-match hints while typing, and an explicit merge.

Independent variables: hints of existing words, whether merge exists, whether suggestions sort by use. Dependent variables: unique word count, cluster size, miss rate retrieving a class with one word, abandonment of tagging.

One lab session will not explode. Use several batches, allow overnight. Do not count machine-applied tags in the human explosion. An admin merge with no preview measures broken trust, not convergence.

Where it stops holding

A private personal library can stay messier; convergence hints should be weaker. Legal vocabularies forbid merging statutory codes. In a multilingual product, Chinese and English for one concept are not a synonym error; do not auto-merge, and offer “filter these language variants together.” Machine keyword floods (an entity per paragraph) are not the human vocabulary to converge; isolate them.

Applying it

• While typing a new word, list similar existing words and default to an existing one; create-new is secondary.
• Offer merge A into B, stating the scope (my vocabulary / this space), with undo.
• Sort the filter list by use; long-tail words remain searchable without filling the first screen.
• Verify: leave two spellings for one purpose, and see whether the second tagging is nudged toward the existing word. After merge, either old word should retrieve the full set. Before merge, people should be able to say what will happen.

Related

• Within the group: H8.10.1 Tag systems need user-defined labels, not only presets · H8.10.2 Multi-tag retrieval must state AND versus OR · H8.10.4 System tags and user tags need a visible source
• Adjacent: H8.05 Save for Later · G1.10 Controlled Vocabularies and Synonyms · H8.13 Content Search and Location
• Search terms: tag proliferation · synonym merge · vocabulary control