Content modelling and knowledge organisation: metadata, vocabularies and taxonomies

Every metadata scheme selects. A field exists because someone decided a distinction mattered. A list of allowed values exists because someone decided which alternatives a form should offer. A hierarchy exists because someone placed one term under another. None of those decisions is recorded in the software, which accepts whatever scheme it is given. This ninth unit turns to what the facts are called. This lecture argues that metadata, controlled vocabularies and taxonomies are claims about which distinctions matter. It argues that a vocabulary earns its keep only when it is tested against how people file and search, not when it mirrors the structure of the organisation that built it.

Two bodies of evidence pull in opposite directions. The standards tradition treats a good vocabulary as a precondition for retrieval, measured by the relations it declares. The tagging literature shows shared order arising at scale with no designer at all.

Organising is a language of description

Svenonius and Arms (2000) built the theoretical foundation. They integrate descriptive cataloguing, subject cataloguing, indexing and classification into one framework in which organising information uses a special language of description, a bibliographic language. Their claim is that a retrieval mechanism is only as effective as the intelligence put into the description it searches. The design work therefore happens inside the vocabulary, before any query is run.

Joudrey and Taylor (2017) give the systematic textbook treatment. It covers the Semantic Web, linked data and newer models such as IFLA-LRM and RiC alongside the historical methods, and a description language that keeps being revised is a design commitment rather than a fixed inheritance.

A metadata field is a decision about which distinction matters

Riley (2017) supplies the definitional account in the NISO primer, separating the types and functions of metadata and asking what metadata is for rather than what it is made of. The useful distinction is between description that helps a person recognise an item, structure that helps software assemble it, and administration that records who may use it and for how long.

Baca (2016) makes the cost explicit: metadata is created, maintained and migrated, so a field nobody uses is not free. Rockley and Cooper (2012) carry the content-modelling half, arguing that content should be modelled as structured elements so one set can be reused across delivery channels. Separating meaning from presentation makes reuse possible, and reuse is what judges a vocabulary.

A vocabulary declares relations, and an ontology is the strongest version

NISO (2010) fixes what a controlled vocabulary is and how its terms relate, with the practical effect that a field offers a finite, closed set of values. ISO (2011) extends the concern to interoperability, treating a thesaurus as something that must work with other vocabularies rather than stand alone. A filter needs a finite set of values to offer, and two systems can exchange descriptions only if they agree on what a term means. The standard is written for one language, so a searcher who knows the thing by another name has no term to choose.

The standard is concrete about what a vocabulary declares. Its glossary fixes three relations between terms: equivalence, written USE with the reciprocal UF, where a non-preferred term points to the preferred one; hierarchy, written BT and NT, for a broader and a narrower term; and association, written RT, for two connected terms (NISO, 2010). A term that carries these relations can be interpreted by a system that never saw the collection it came from. A term that carries none is a string, and that is the difference the standard means by control.

An uncontrolled field is cheap to populate and expensive to query, because every writer invents a different label for the same thing. A controlled one moves the cost to design, where it is paid once and then maintained.

A photograph with only a free-text note, taken in Cairo in 1900 by an unknown photographer, is findable only by the words a searcher happens to guess. Given a declared date, a place term from a published list and a creator field that can be marked unknown, the same photograph can be filtered, sorted and linked to others from the same place, and only the second description can be acted on by software the catalogue did not build.

Gruber (1993) gives the technical sense of an ontology as a specification of a representational vocabulary for a shared domain, covering classes, relations and functions. He describes a mechanism for defining ontologies that are portable across representation systems. Noy and McGuinness (2001) reduce the task to decisions a team can take, defining classes, properties and instances in a worked example.

What the strongest version cannot supply is an answer to whose distinctions are specified, the question the next section raises.

Tagging produces stability without declaring anything

Mathes (2004) named the phenomenon, describing folksonomies as cooperative classification through shared metadata, in which users tag resources with their own words, and setting out its strengths and weaknesses against professional classification. Guy and Tonkin (2006) take the balanced position. Tags are no replacement for formal systems, and that is precisely what makes tagging useful, because it captures distinctions a formal scheme would never anticipate.

Two papers test whether a shared vocabulary can emerge without design. Golder and Huberman (2006) analysed collaborative tagging and reported regularities in user activity, tag frequencies and the kinds of tags used, together with a marked stability in the relative proportions of tags within a given resource. Halpin, Robu and Shepherd (2007) asked the sharper question, whether coherent categorisation can emerge from unsupervised tagging. They found that the distribution of tag frequency at popular bookmarks followed a power law, and they built a generative model from it. Neither shows that emergent tags declare the relations a thesaurus provides. Both show that use alone produces order of a sort.

Metadata that machines can act on

Wilkinson and colleagues (2016) give the thesis a measurable standard. The FAIR principles are a concise and measurable set of guidelines for scientific data management. Their distinguishing emphasis is that machines should be able to find and use data automatically, in addition to supporting reuse by people. That reframes the test of a vocabulary: the question is no longer whether it captures the domain faithfully but whether its descriptions can be acted on by software that did not build them.

A vocabulary that fails the reuse test fails quietly: nothing in the interface reports that a field meant something different to the system next door.

Contested question: does order have to be designed, or does it emerge?

The literature disagrees, partly about what counts as order.

The first position holds that a usable vocabulary must be designed. Svenonius and Arms (2000) treat the intelligence put into the description as the ceiling on what retrieval can achieve. NISO (2010) and ISO (2011) make the relations among terms explicit, so a term can be interpreted outside the system that issued it. Gruber (1993) and Noy and McGuinness (2001) treat the vocabulary as a specification that must exist before it can be shared. On this account emergent tags fail not because they are messy but because they declare nothing, so no second system can act on them.

The second position holds that order can emerge. Mathes (2004) and Guy and Tonkin (2006) argue that the looseness of tags is their advantage, since users describe items in their own terms at no cataloguing cost. Golder and Huberman (2006) found stability in the relative proportions of tags attached to a resource and modelled it as imitation. Halpin, Robu and Shepherd (2007) found a power-law structure in tag distributions and built a generative account of it. On this account the stability is real and needs no designer, so hand-building a taxonomy is avoidable work.

This lecture takes the position that the two findings are about different quantities. Golder and Huberman (2006) and Halpin, Robu and Shepherd (2007) measure whether a distribution stabilises within one service, an observational result. Svenonius and Arms (2000) and the standards measure whether a term carries a declared relation another system can consume, a design requirement about interoperability. A service can have the first without the second. What would settle it is an experiment that describes one corpus both ways and reports both outcomes together. The first outcome is whether people find items using their own terms. The second is whether a separate system can consume the descriptions without a human in the loop. Each camp measures only the outcome favourable to its position, which is why the disagreement has not closed.

Rate each side by its evidence. Svenonius and Arms (2000), NISO (2010) and ISO (2011) state a design requirement, and Gruber (1993) and Noy and McGuinness (2001) construct specifications, so the design side rests on theory and standards rather than measurement. Golder and Huberman (2006) and Halpin, Robu and Shepherd (2007) report observation inside one service rather than a comparison between systems. The strongest objection is that a settled distribution is not a declared relation, so the stability both studies report may be evidence of use rather than of the order a second system needs. While the dispute is open, a designer tests whether the descriptions can be consumed without a human in the loop.

Where the argument is heading

Two named lines point forward. The first is the revisable description language. Joudrey and Taylor (2017) document newer models such as IFLA-LRM and RiC replacing earlier conventions, and they concede that a vocabulary is a settlement rather than a fact about the world. The second is measurable, machine-actionable metadata: Wilkinson and colleagues (2016) turn reuse from an aspiration into a property a pipeline can test, and Akhtar and colleagues (2024) give a worked format, Croissant, that describes a dataset's resources, structure and default machine-learning semantics in one file and has been adopted by the large dataset repositories.

The sources this unit rests on

• Akhtar, M., Benjelloun, O., Conforti, C. et al. (2024). Presents Croissant, a metadata format describing a dataset's resources, structure and default machine-learning semantics in one JSON-LD file built on schema.org, and reports human raters judging it readable, complete and concise. It is a format proposal with a readability check rather than a retrieval study.
• Baca, M. (ed.) (2016). Introduction to Metadata. Getty Research Institute. The Getty's practitioner introduction to metadata; only the publication's own page was verified. It treats metadata as an ongoing cost of creation and maintenance rather than a one-off task.
• Golder, S. A. and Huberman, B. A. (2006). Usage patterns of collaborative tagging systems. Journal of Information Science. The authors analysed collaborative tagging and found regularities in user activity and tag frequencies, plus a marked stability in the relative proportions of tags within a resource, modelled as imitation. The stability holds within one service.
• Gruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition. Gruber defines an ontology as a specification of a representational vocabulary for a shared domain and makes such definitions portable across representation systems. It gives the strongest version of a vocabulary as an explicit shared specification.
• Guy, M. and Tonkin, E. (2006). Folksonomies: tidying up tags? D-Lib Magazine. The authors examine what makes folksonomies work and argue that tags are no replacement for formal systems, but that this is what makes them useful. It frames the contested question from the middle.
• Halpin, H., Robu, V. and Shepherd, H. (2007). The complex dynamics of collaborative tagging. Proceedings of the 16th international conference on World Wide Web. Using social bookmarking data, the authors asked whether coherent categorisation can emerge from unsupervised tagging and found tag-frequency distributions following a power law, the precise form of the emergent-order claim. It tests one service rather than cross-system sharing.
• ISO (2011). Information and documentation: thesauri and interoperability with other vocabularies, part 1 (ISO 25964-1:2011). ISO. The ISO standard on thesauri and their interoperability with other vocabularies; only the bibliographic record was verified. It defines the interoperability requirement the emergent camp does not measure.
• Joudrey, D. N. and Taylor, A. G. (2017). The Organisation of Information. Libraries Unlimited. The book's own title spells the term in the American form, and is rendered here in the house British spelling. The standard textbook of information organisation, covering the Semantic Web, linked data and newer models such as IFLA-LRM and RiC; only the publisher description was verified.
• Mathes, A. (2004). Folksonomies: cooperative classification and communication through shared metadata. University of Illinois Urbana-Champaign (technical report). Mathes describes folksonomies as cooperative classification through shared metadata and reviews its strengths and weaknesses against professional classification; only the report's own page was verified. It names the phenomenon the unit tests.
• NISO (2010). Guidelines for the Construction, Format, and Management of Monolingual Controlled Vocabularies (ANSI/NISO Z39.19-2005). ANSI/NISO. The ANSI/NISO standard on monolingual controlled vocabularies; only the bibliographic record was verified. It fixes what a controlled vocabulary is and how its terms relate.
• Noy, N. F. and McGuinness, D. L. (2001). Ontology development 101: a guide to creating your first ontology. Stanford Knowledge Systems Laboratory Technical Report KSL-01-05. A practical guide to defining classes, properties and instances with a worked example; only the report's own page was verified.
• Riley, J. (2017). Understanding Metadata: What is Metadata, and What is it For? NISO. The NISO primer defining what metadata is and surveying its types and functions; only its own page was verified.
• Rockley, A. and Cooper, C. (2012). Managing Enterprise Content: A Unified Content Strategy. New Riders. The practitioner reference for structured content and a unified content strategy, in which structured elements are reused across delivery channels; only the publisher page was verified.
• Svenonius, E. and Arms, W. Y. (2000). The Intellectual Foundation of Information Organisation. MIT Press. The book's own title spells the term in the American form, and is rendered here in the house British spelling. Svenonius integrates descriptive cataloguing, subject cataloguing, indexing and classification into one framework in which organising information is the use of a special language of description; only the bibliographic record was verified.
• Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J. et al. (2016). The FAIR guiding principles for scientific data management and stewardship. Scientific Data. The authors present the FAIR principles, a concise and measurable set of guidelines whose emphasis is that machines should be able to find and use data automatically so data can be reused.