Data Catalog Vocabulary (DCAT)

DCAT supplies a model for publishing metadata about data catalogs, datasets, and distributions. It helps a reader find what data exists, who publishes it, and which files or services offer access. A dataset and its distributions are different things: one conceptual dataset can be available in multiple formats or endpoints. That distinction supports catalog integration without treating every download as a separate subject.

DCAT metadata does not guarantee that a distribution URL works, that a dataset is complete, or that its licence permits every use. Publisher descriptions can become stale as files and access conditions change. Keep retrieval times, version information, licences, and checks of the actual distribution separately. A catalog entry is a discovery record, not a retained copy of the dataset.

A fictional agency lists a dataset of inspection results. The DCAT record gives the title, the publisher, an update frequency, a licence and two distributions: a CSV file and a query endpoint. A researcher harvests the record and uses the CSV.

A reviewer later asks which version of the data underlies a published figure. The harvested record shows only the current description. The researcher's own record holds the dataset identifier, the distribution URL, the retrieval date, the file's checksum and a retained copy, which together answer the question. Without the checksum and the copy, the best that can be said is that the figure was derived from whatever the URL served on a certain day. The catalog entry helped to find the data and cannot reconstruct it.

A catalog entry is not a copy of the data and is not an assessment of its quality. It is a discovery record. Finding a dataset in a catalog does not show that it is complete, accurate or current.

DCAT is also not a provenance model. It has a few properties for source and version, but a full account of how a dataset was produced calls for a provenance vocabulary, and DCAT is designed to be used together with one. Finally, a dataset and a distribution are different objects, and counting distributions as datasets overstates how much data a catalog holds.

Catalog descriptions go stale in ordinary ways. Datasets are updated without the description being touched, distributions move to new URLs, licences change and publishers merge. Harvesting systems typically refresh on a schedule, and a harvested record is a copy of a description at a time. When a record matters as evidence, store the harvest date with it and the identifier of the harvested source, so that a later reader can distinguish what the publisher said at the time from what the publisher says now. A catalog that records its own modification dates and the dates of the underlying data gives reviewers a starting point, and both dates should be checked against the data rather than accepted.