@inproceedings{kovacs-etal-2026-twentys,
    title = "Twenty{'}s Plenty: Semantic Scaffolding and Span Architecture for 19-Label {NER} in Medieval {L}atin Charters",
    author = "Kov{\'a}cs, Tam{\'a}s  and
      Consolo, Giuseppe  and
      Vogeler, Georg",
    editor = {Hamilton, Sil  and
      {\"O}hman, Emily  and
      Hicke, Rebecca M. M.  and
      Bizzoni, Yuri  and
      Bax, Axel  and
      Matthews, Jacob A.  and
      H{\"a}m{\"a}l{\"a}inen, Mika},
    booktitle = "Proceedings of the 6th International Conference on Natural Language Processing for the Digital Humanities",
    month = jul,
    year = "2026",
    address = "San Diego, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.nlp4dh-1.22/",
    pages = "236--241",
    ISBN = "979-8-89176-427-9",
    abstract = "This study investigates whether a high-quality, 19-label named entity recogniser for medieval Latin charters can be constructed using only a few hundred annotated sentences. The authors introduce ``semantic scaffolding,'' an innovation that utilizes richly descriptive English label phrases as prompts to activate latent multilingual knowledge within the model. This is paired with a custom span-based architecture utilizing XLM-ROBERTa-large, 4-head attention pooling to handle long property descriptions, and a hybrid loss system including Asymmetric Focal-Dice and InfoNCE contrastive terms. Results demonstrate that semantic scaffolding enables fine-tuned GLiNER to reach 80.8{\%} overlap F1, while the custom architecture achieves 83.4{\%} overlap F1 using only 298 training sentences. Significantly, the paper provides an empirical demonstration that domain-specific pre-training on medieval Latin offers no performance advantage once task-specific fine-tuning is applied. While the model excels at frequent categories like PER (95.7{\%} F1) and LOC (93.5{\%} F1), challenges persist for rare, position-dependent legal categories such as LEG (53.1{\%} F1) and TRANS (52.6{\%} F1)."
}