Content tagging tools suggest labels that help people organize and find information. Begin with the purpose of the tags and a small reviewed sample. Evaluate suggestions against that purpose before allowing them to drive navigation, reporting or other consequential actions.
Separate classification from page metadata
A subject tag such as ACCESS describes a topic. An HTML description summarizes a page for tools that may use it. A search-result snippet is the text a search service chooses to display. These serve different purposes and should have separate owners and rules.
MetaTagger 3.5, announced in 2003, addressed enterprise classification and taxonomy management. This guide follows that task: assigning topical labels consistently. Generating a large list of keywords alone provides little evidence that readers can find the right item.
Define a reader task for each category: ACCESS might help staff find guidance about entering a building or obtaining an accommodation. Decide whether a page can have several labels and whether a label needs review before use.
Give terms stable identities and useful boundaries
| Code and preferred label | Alternative wording | Boundary |
|---|---|---|
| ACCESS — Access | Entry; accessibility | Building access and accommodations; exclude account passwords. |
| VISITS — Service visits | Appointments; attendance | Preparing for or attending a service visit. |
| ACCOUNTS — User accounts | Login; password | Account access and recovery; exclude physical building entry. |
Record examples and counterexamples with the definition. “Access problem” is ambiguous until context establishes whether it concerns a building or an account. Have reviewers resolve disputed examples before treating their labels as an evaluation reference.
W3C SKOS provides concepts, preferred and alternative labels, relationships and mappings for knowledge-organization systems. A simple local vocabulary can use these distinctions without claiming that an ordinary spreadsheet is a complete SKOS implementation.
Measure both extra tags and missed tags
Download the 20-row evaluation CSV and instructions and vocabulary worksheet. Keep the reference label and the tool's proposal in separate columns. Calculate each category separately; a common easy category can hide poor performance on an important rare one.
If a denominator is zero, report that the metric is undefined for that sample instead of substituting a success value. Keep the sample size, selection method and disputed labels beside every result.
Choose a review rule from the errors
Examine false positives and missed items by cause: ambiguous wording, missing context, a vocabulary gap or a poor extraction. Rules may handle well-defined repeated phrases, while statistical or AI-assisted tools may suggest broader matches. Compare them on the same held-out examples after configuration is complete.
A confidence number is a model output, not a guarantee. Evaluate any proposed threshold with reviewed data and the consequences of a mistake. Sensitive routing, permissions or records actions need their own approved controls; a topical suggestion should not silently become authorization.
Give editors a way to accept, reject or change a suggestion and explain recurring errors. Retain the tool/configuration version and vocabulary version so later comparisons mean something.
Keep the vocabulary and test set usable
Assign an owner to new terms, merges and retirements. Preserve stable codes when a label merely changes wording; map old codes when the concept changes. Check dependent filters, reports and integrations before releasing an update.
Add representative failure cases to the regression set while keeping a separate sample for fresh evaluation. Review reader search failures alongside tagging metrics. A mathematically better classifier may still produce categories that readers do not understand. Use the metadata mapping guide when tags move between systems.