mschonhardt/ius-commune-citations-ner

Citations and named entities in Latin texts of the ius commune

A Flair sequence tagger that finds allegations of the learned law and the named entities in Latin legal prose of the 13th to 16th centuries. It marks citations of the Corpus iuris civilis and the Corpus iuris canonici in their abbreviated medieval form ("ff. de pac. l. tale pactum. §. fi.", "extra de proba. c. cum causam."), citations of the Libri feudorum and other works, the jurists and glosses invoked by siglum ("Bar.", "Inno.", "Io. And.", "gl. ord."), and statutes, persons, places, institutions and dates.

The model was built for the machine-assisted study of the legal cultures of the late Middle Ages and the early modern period. Once the citations and names are marked, a corpus of consilia, commentaries or tractatus can be indexed by the authorities it cites, the jurists it follows and the parties and places it concerns, and the argumentative texture of the ius commune becomes searchable at scale. The tagger annotates the texts of the Consilia portal.

Developed in the project Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL Mainz), funded by HMWK Hessen, PI Prof. Dr. Michael Schonhardt.

Labels

label	meaning	examples as printed
CIT_CIV	citation of the Corpus iuris civilis (Digest, Codex, Institutes, Authenticum and Novels)	ff. de ver. ob. l. 1., C. de testa. l. hac consultissima, insti. de inoffi. te. §. mater, in Auth. vt cum de appe. cogno.
CIT_CAN	citation of the Corpus iuris canonici (Decretum, Liber Extra, Sextus, Clementines)	extra de proba. c. cum causam., xxx. q. v. c. nostrates, Clem. vnica de testib.
CIT_FEUD	citation of the Libri feudorum	in vsib. feu. ti. de capi. qui cur. vend.
CIT_OTHER	other cited work (Scripture, philosophy, literature)	Matth. xix., Arist. in v. Ethico.
JURIST	jurist or gloss cited by name or siglum	Bar., Inno., Cyn., Io. And., glo. ord.
STATUTE	statute or local law referred to	statuto communis Perusii, sub rub. quod mulieres habentes
PERSON	party or other person of the case	Nicolina, dominus Ioannes de Angelo
PLACE	place	castrum Sauigliani, Paduę
ORG	institution, commune or church	commune Eug., communitas Senarum
DATE	date	anno domini 1388., Anno Domini Millesi. 383. indictione 6.

A span is the whole reference as the print gives it, from the first siglum to the last element of the same reference, while chains of references joined by "et" are separate spans. Spelling and abbreviation follow the print.

Architecture

Flair SequenceTagger with a CRF over stacked embeddings:

• mschonhardt/MediBerta, a continued pre-training of FacebookAI/xlm-roberta-large on medieval Latin, fine-tuned during training with first-subtoken pooling and document context,
• mschonhardt/latin-legal-forward and mschonhardt/latin-legal-backward, contextual string embeddings (Flair character language models trained on Latin legal text), kept frozen.

The tags are in BIOES format, with 10 span types and 43 tags.

Training data

The training data are made from the texts of the Consilia edition: Baldus de Ubaldis, Consilia (Venice 1575, five volumes, 2,514 consilia) and Nicolaus de Tudeschis (Panormitanus), Consilia (Venice 1571, 226 consilia). All 10,322 text units of the edition were annotated by Gemini 4 Argon, which was asked to mark every citation and name as a verbatim quotation. A check of a stratified sample of 150 consilia by reading found that this annotation misses hardly any citation (recall about 99.9 % in the sample). The quotations were located in the text and converted to BIOES tags. The split is by consilium, so that no document appears in two sets.

split	sentences	tokens	spans
train	89,250	2,358,486	91,088
dev	11,990	307,665	11,689
test	10,511	264,367	9,777

Spans per label in the training set: CIT_CIV 42,546, PERSON 17,302, JURIST 13,907, CIT_CAN 11,101, PLACE 2,360, ORG 1,699, STATUTE 802, CIT_FEUD 786, DATE 295, CIT_OTHER 290.

Training

The model was trained for 20 epochs with a learning rate of 5e-5 (linear schedule with 10 % warm-up), mini-batches of 16 sentences, shuffling and mixed precision, and a checkpoint after every epoch. The best model was chosen by the span F1 on the dev set. Training took 5.1 hours on one GPU on 29 October 2026, about 15 minutes per epoch. The dev F1 (micro) rose from 0.850 after the first epoch to 0.912 after the last, which was also the best. The training log (training.log) and the losses and dev scores per epoch (loss.tsv) are in this repository.

Evaluation

The scores are span-level, with an exact match of boundaries and label, on the held-out test split (10,511 sentences, 9,777 spans).

label	precision	recall	F1	support
CIT_CIV	0.950	0.953	0.952	4,861
PERSON	0.903	0.903	0.903	1,871
JURIST	0.932	0.929	0.930	1,447
CIT_CAN	0.881	0.890	0.885	915
PLACE	0.636	0.643	0.640	272
ORG	0.727	0.747	0.737	178
STATUTE	0.579	0.619	0.598	118
CIT_FEUD	0.820	0.820	0.820	50
DATE	0.625	0.714	0.667	35
CIT_OTHER	0.364	0.267	0.308	30
micro avg	0.911	0.914	0.912	9,777
macro avg	0.742	0.748	0.744	9,777

The citations of the two corpora, the jurists and the persons, which make up most spans, are found reliably. The rare types (statutes, dates, other works) have few training examples and are less reliable.

Use

from flair.models import SequenceTagger
from flair.data import Sentence

tagger = SequenceTagger.load("mschonhardt/ius-commune-citations-ner")
s = Sentence("Item exclusa filia videtur excludi omnis descendens per filiam. ar. ff. de colla. bo. l. dedit dotem. "
             "et C. de nat. lib. l. fi. vt eleganter no. Bar. in l. 1. ff. vti possi.")
tagger.predict(s)
for span in s.get_spans("ner"):
    print(span.tag, round(span.score, 2), span.text)

tag_text.py in this repository tags a text file and returns the spans with character offsets. The input should be tokenised like the training data (runs of letters or digits and single punctuation marks, which Flair's default tokenizer comes close to), with the abbreviations as printed. The model expects the transcription conventions of the Consilia edition, that is "et" for the ampersand, "s" for the long s, "ae" for the ligature and abbreviation marks expanded (nõ becomes non).

Limitations

The model was trained on two authors and two prints (Venice 1571 and 1575, roman type, consilia). Other genres such as commentaries, glosses or statutes and other orthographies need a check and probably further fine-tuning. Nested references, for example a gloss on a lex, form one span of the outer type.

Environmental impact

The following is an estimate and not a measurement, since no power meter was attached.

step	GPU time	assumed draw	energy incl. host and cooling (PUE 1.4)
training (20 epochs)	5.1 h	400 W average on a 600 W card	about 2.9 kWh
tagging the corpus (10,322 units)	4 min	400 W	under 0.1 kWh

The hardware was one NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB) in a university server in Darmstadt, Germany. With the German grid average of about 0.38 kg CO2e per kWh (Umweltbundesamt, 2024 mix) the training amounts to roughly 1.1 kg CO2e. The estimate follows the method of the Machine Learning Impact calculator (Lacoste et al. 2019), multiplying hours, power and carbon intensity with a PUE factor. The silver annotation by Gemini 4 Argon ran as a cloud service, whose energy use is not published and is not included here.

Citation

Schonhardt, Michael: ius-commune-citations-ner. Citations and named entities in Latin texts of the ius commune (Flair sequence tagger). Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL Mainz), 2026. https://huggingface.co/mschonhardt/ius-commune-citations-ner

@misc{schonhardt2026iuscommune,
  author       = {Schonhardt, Michael},
  title        = {ius-commune-citations-ner: Citations and named entities in Latin texts of the ius commune (Flair sequence tagger)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/mschonhardt/ius-commune-citations-ner}},
  note         = {Developed in the project Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL Mainz), funded by HMWK Hessen}
}

Related

• The Consilia portal: https://dekm-tud.github.io/coding-consilia/ , code and data: https://github.com/dekm-tud/coding-consilia
• kraken models for the same prints: mschonhardt/consilia-baldus-venice1575-kraken-seg and -kraken-rec
• PageXML ground truth: mschonhardt/consilia-baldus-venice1575-pagexml