BibNER-Werk: Bibliographic NER for Historical German Catalogs

General-purpose named entity recognition does not identify the title and responsible person needed to link historical German bibliographic catalog strings in the Deutsche Digitale Bibliothek (DDB) to GND1 Work authorities, since the task’s label set and centuriesspanning historical register have no counterpart in standard training corpora. This paper uses ISBD punctuation as a tiered silverlabeling signal and evaluates a deterministic rule-based baseline alongside zero-shot, few-shot, and fine-tuned GLiNER-multi and fine-tuned XLM-R-large on a stratified, human-annotated gold set. The rule-based baseline beats both fine-tuned systems on untrained data; on trained data, only GLiNER-multi’s fine-tuning gains a real, confirmed precision advantage, and XLM-R-large’s never surpasses the rule