Elamur Research⌕
☰
home / papers / 2026 / Language Models for the Restoration of Latin Legal Manuscripts

Language Models for the Restoration of Latin Legal Manuscripts

Summary

Many medieval notarial deeds from Bologna are physically damaged, and the gaps (lacunae) hide historical information. The authors treat restoration as masked-token infilling with encoder PLMs, assuming the number of missing characters can be estimated from the size of the damage. The main obstacle is that a character-length gap does not map cleanly onto BPE subword tokens, and damage often cuts words at arbitrary points. To handle this, they built a four-stage decoding pipeline:

  1. Expand the lacuna to whole-word boundaries, keeping the cut-off fragments as a known prefix and suffix.
  2. Generate several mask-count hypotheses from token-length statistics in the training data.
  3. Run iterative beam search with a hard prefix constraint and a suffix reward.
  4. Pick the best-scoring candidate.

They fine-tuned the Latin-only LaBERTa and the multilingual PhilBERTa on Latin deeds written by 254 notaries. Both models were compared against simpler decoding baselines, and results were broken down by document section. The pipeline greatly outperforms the baselines, and the monolingual model is generally stronger. Exact recovery is still far from reliable, so the system works best as a suggestion tool for paleographers.

Key points

  • Best results (LaBERTa with the full "Iterative Boundary" decoding):
    • 5–10 character gaps: HR@1 35.47%, HR@10 56.09%, CER 55.40.
    • 10–20 character gaps: HR@1 18.75%.
  • Baselines: iterative decoding without lacuna expansion and simple decoding reach only about 1–2% HR@1, with CER around 72.
  • Monolingual vs. multilingual: PhilBERTa (Latin, Ancient Greek, English) trails LaBERTa overall, with 28.56% vs. 35.47% HR@1 on short spans.
  • Results by section: formulaic protocol sections are easiest (about 41% HR@1). Rogationes, the notary's summary notes, are hardest (about 16%) and the most lexically varied (TTR 0.310, 19.7% hapax). On rogationes PhilBERTa edges ahead slightly, but the difference is not statistically significant.
  • Setup: 8:1:1 data split, documents averaging about 380 words, and five independent fine-tuned checkpoints per model, reported as mean ± 95% CI.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages · Natural language processing · Law · History & humanities

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/5g8oqsnzp537

@inproceedings{zhang2026language,
  title = {{Language Models for the Restoration of Latin Legal Manuscripts}},
  author = {Zhang, Shibingfeng and Caraffa, Edoardo and Zuffrano, Annafelicia and Modesti, Maddalena and Colavizza, Giovanni},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/5g8oqsnzp537},
  url = {https://doi.org/10.63317/5g8oqsnzp537}
}

References

2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yannis Assael, Thea Sommerschield, Alison Cooley, Brendan Shillingford, John Pavlopoulos, Priyanka Suresh, Bailey Herms, Justin Grayston, Benjamin Maynard, Nicholas Dietrich, et al. 2025. Contextualizing ancient texts with generative neural networks. Nature, 645(8079):141–147. find
  2. Giorgio Cencetti. 1977. Notariato medievale bolognese. Tomo I: Scritti di Giorgio Cencetti. Number 3, 1 in Studi storici sul notariato italiano. Consiglio nazionale del notariato, Roma. find

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-language-models-restoration-latin-legal/