Twenty's Plenty – Medieval Latin Span-NER

Named entity recognition for Medieval Latin charters with 19 fine-grained diplomatic entity types (persons, titles, places, institutions, legal clauses, dates, money, measures, ...).

pip install torch transformers huggingface_hub
pip install jupyter matplotlib pandas   # only for the notebook

The first run downloads the weights (~4.5 GB, cached afterwards) plus the tokenizers of FacebookAI/xlm-roberta-large and BAAI/bge-m3. A GPU is used automatically if available; CPU works but is slower.

python medieval_latin_ner.py PannHActa_1416_VIII_10_charter.txt                  # table on stdout
python medieval_latin_ner.py charter.txt --json entities.json --html view.html  # save results
python medieval_latin_ner.py charter.txt --labels PER ACTOR LOC INS             # filter labels
python medieval_latin_ner.py --text "Ego Heinricus comes de Hals ..."           # inline text

Options: --threshold (default 0.5), --flat (no nested entities), --device cpu|cuda.

from medieval_latin_ner import MedievalLatinNER

ner = MedievalLatinNER()
entities = ner.predict("Ego Heinricus comes de Hals dedi ecclesie Sancti Petri ...", threshold=0.5)
for e in entities:
    print(e["label"], e["score"], e["text"], e["start_char"], e["end_char"])

Every entity is a dict with label, text, score, start_char, end_char, start_word, end_word. By default entities may be nested (e.g. a PER inside an ACTOR), as in the paper's annotation scheme.

jupyter notebook medieval_latin_ner_demo.ipynb   # spaCy-style highlighted charter, label table, statistics

The 19 entity types, grouped into the four documentary layers of the annotation scheme. The English descriptions are the label prompts used by the model (semantic scaffolding, see the paper); they are also available as ner.label_descriptions.

A custom bi-encoder span model: XLM-RoBERTa-large encodes the text, spans (up to 80 tokens) are represented by start/end tokens, multi-head attention pooling and a width embedding, and scored against frozen BGE-M3 embeddings of English label descriptions ("semantic scaffolding"). Training used an asymmetric focal + Dice loss with InfoNCE and hard-negative mining. Because of this custom architecture the standard pipeline() API does not work – use medieval_latin_ner.py, which loads the architecture file from the model repository.

MedievalLatinNER.predict() splits long charters into overlapping windows (window=150, overlap=40 tokens) and merges the results. This is necessary because the chunking in the repository's span_ner_model.py counts 512 words, while the XLM-R tokenizer truncates at 512 subwords: with the raw SpanNERModel.predict(), everything after roughly the first 250 words of a charter stays untagged. (The charter has 460 tokens: the raw call stopped at character 1818 of 3186.)

The training data are 13th-century Austrian and Central European charters. Other periods, chanceries and orthographies (such as the 15th-century royal Hungarian mandate in the demo) can give lower accuracy, particularly for rare labels (PROP, EST, TAX, ...). Check the results before using them in research.

@inproceedings{kovacs2026twenty,
  title     = {Twenty's Plenty: Semantic Scaffolding and Span Architecture for 19-Label NER in Medieval Latin Charters},
  author    = {Kovács, Tamás and Consolo, Giuseppe and Vogeler, Georg},
  booktitle = {Proceedings of the 6th International Conference on Natural Language Processing for the Digital Humanities},
  pages     = {236--241},
  year      = {2026}
}

@dataset{consolo2026ner,
  title     = {Named Entity Recognition Dataset for Medieval Latin Charters},
  author    = {Consolo, Giuseppe and Kovács, Tamás and Vogeler, Georg},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19009431}
}