Plan a pilot box
Menu

AI-assisted cataloging with human review

Models draft names, places, dates, subjects and descriptions from your text and images. A cataloger reviews the suggestions, and every field says whether a person has checked it.

Most collections have a cataloging backlog that no budget will clear by hand: boxes listed as “family papers, 1890s” with nothing about who wrote to whom or what happened. Language models can read recognized text and images and draft the fields a cataloger would fill in. They are also wrong often enough that nobody should publish their output unchecked. We use them to draft, people to decide, and labels to show which is which.

Why every suggestion is reviewed

The Library of Congress tested machine-generated catalog data in its Exploring Computational Description experiments (2022 to 2024). Identifiers, authors and titles were extracted at around 85 percent accuracy, but subject headings, genre and dates were generated accurately “at or under 50% of the time”. Its conclusion: “no current AI tools are currently good enough to run fully automatically for this task.” A pilot box measures the rate on your own records, but the starting assumption is the same. A model’s suggestion is a draft.

What the models suggest

Field Where the suggestion comes from What the reviewer checks
Names Signatures, salutations, text, photo backs Right person, authority form, no invented people
Places Datelines, postmarks, studio imprints, text Right place, authority form, current or historical name
Dates Datelines, postmarks, ledger entries, mount styles Exact or approximate, and how it was inferred
Subjects The content of the text or image Fit with your vocabulary, nothing sensitive mislabeled
Description A short summary drafted from the text Accuracy, tone, names spelled as in the authority

Rights statements, access restrictions and descriptions of culturally sensitive material are written by people, not suggested by models.

Matching to your vocabularies

Suggestions are matched to the vocabularies your catalog already uses. For subjects that is often Library of Congress Subject Headings, which the Library describes as actively maintained since 1898. For names and places it is often your own authority lists: the families, businesses, schools and place names that recur in your collection. Where a suggested term is not in the vocabulary, it is proposed for your approval, never added silently. The metadata mapping worksheet records which vocabulary governs each field.

Every machine field stays labeled

Each descriptive field carries a provenance value:

  • Machine suggestion: drafted by a model and not yet reviewed.
  • Reviewed: checked and accepted by a person.
  • Edited by reviewer: changed by a person after review.
  • Human-created: written by a person from the start.

The reviewer and date are recorded with each value. The labels travel with the record: as a column in CSV exports, and as a note in Dublin Core, MARC and EAD exports, so your system and your users can still tell them apart after import. No professional rule requires this; it is our method, set out in our AI policy.

Example (fictional). Letter ACHS-3-019 from the Aster County Historical Society’s Box 3, Ida Okerlund to her sister Marit, dated February 3, 1894, reads in part: “The smelter men walked out on Monday and Walter says the store will carry them on credit until the matter is settled.”

Field Value Label
Names Okerlund, Ida; Okerlund, Walter Reviewed
Places Wren Gulch (Colo.) Reviewed
Subjects Labor disputes; Credit Machine suggestion

The names and place have been checked against the society’s authority list. The subjects are still drafts, and the record says so until a cataloger reviews them. The fictional example collection shows more records like this one.

The review workflow

  1. Agree the map. Which fields, which vocabularies, which export format, and which fields must be reviewed before delivery.
  2. Machine pass. Models read the recognized text from text recognition and the images, and draft suggestions.
  3. Review queue. A cataloger works through the suggestions field by field, with the page image and the text side by side. Accept, edit or reject.
  4. Second check. A sample of reviewed records is checked again, and the error rate is reported per field.
  5. Export. Records leave with their labels intact.

Review can be done by our catalogers, by your staff in the same queue, or split between the two. Fields you choose not to review before delivery are delivered labeled as machine suggestions.

Exports for your collection system

  • CSV mapped to your system’s import template, for example Omeka’s CSV Import or CollectiveAccess’s CSV import.
  • Dublin Core, the 15-element set, with DCMI’s recommended dcterms namespace where your system accepts it.
  • MARC 21, as MARCXML, for library catalogs and for ArchivesSpace.
  • EAD finding aids in the version your system imports: EAD4 is now current, the official EAD site still provides EAD3, and EAD 2002 is deprecated but still widely used. Archival description follows DACS.

Your images and text are not used to train a third party’s general model without your written permission. The Reading Room’s cataloging section covers vocabularies, provenance labels and exports in more depth, and the records can feed collection search directly.

Send one box first.

A pilot box shows you the images, the recognized text and the catalog records before you commit a whole collection. Tell us what is in the box and what you want people to find.

See labeled records in a pilot box

Studio and contacts

24 South Weber Street, Suite 225
Colorado Springs, CO 80903
(719) 244-8450sales@dastok.com