homologia

OIKONOMIA-Homologia — Relation Extraction for Greek Documentary Papyri

ὁμολογία homología, "the acknowledgment" — the contract formula that binds parties to a deal. It is the most common transaction word in the corpus this model was trained on (ὁμολογῶ and its forms, ~3,000 attestations).

A span-pair relation-extraction model that turns tagged entities in a documentary papyrus into the structure of a transaction: who is a party to the deal, who paid whom, which amount prices which commodity, which measure attaches to which quantity, which tax a payment discharges.

It is the relation arm of OIKONOMIA, a project turning the ~68,000 Duke Databank (DDbDP) papyri into a structured, auditable database of ancient economic life. It runs on top of OIKONOMIA-Grammateus — Grammateus finds the spans, Homologia links them. At corpus scale the pair produced 228,945 relations over 61,249 documents, including 16,315 PARTY_OF edges, which became the published OIKONOMIA-DB.

⚠️ This is not a transformers AutoModel. The head is custom, so the checkpoint is a PyTorch state_dict (relation_head.pt) plus a config.json describing how to rebuild it. See How to Get Started — you need the OIKONOMIA package, not just from_pretrained.

---

Model Details

Model Description

• Developed by: Abderahmane Ainouche (OIKONOMIA project)
• Model type: SpERT-style span-pair relation classifier over a RoBERTa encoder
• Language: Ancient Greek (grc) — documentary, not literary
• Licence: apache-2.0, inherited from the encoder
• Encoder: bowphs/GreBerta + domain-adaptive pretraining on the papyri corpus
• Parameters: 129.1M · Context: 512 tokens
• Output space: 12 classes (11 relations + NO_RELATION) over 13 entity endpoint types

Architecture

Per candidate entity pair the head sees:

• max-pooled head and tail span representations,
• the text between the two spans,
• a wide left context reaching back before the payer, where the direction verb lives in Greek contract formulae,
• the CLS state of the window,
• learned entity-type embeddings (type_dim 64) and pair features (feat_dim 16),

with dropout 0.2, and decoding masked by a schema constraint (which entity-type pairs may hold which relation).

Model Sources

• Repository: https://github.com/abderahmane-ai/oikonomia
• Sibling model: OIKONOMIA-Grammateus (entities)
• Dataset produced with it: OIKONOMIA-DB

---

Uses

Direct Use

Structured information extraction over documentary Ancient Greek — contracts, receipts, leases, loans, sales, tax payments — to populate an economic database. Built for research in papyrology, digital humanities and ancient economic history.

Out-of-Scope Use

• Standalone use. It classifies pairs of given spans. Without an entity model it does nothing.
• Payment direction as fact. PAID_BY scores F1 0.145. Treat a predicted direction as a hypothesis; we exclude it from the published database entirely.
• Literary Greek, epigraphy, or Modern Greek.
• Dense registers. Candidate generation is quadratic in entity count; giant lists (hundreds of entities) need a cap. The corpus run skips 35 such documents by design.

---

How to Get Started

pip install git+https://github.com/abderahmane-ai/oikonomia

from oikonomia.relations.model import load_homologia

model, cfg = load_homologia("ainouche-abderahmane/homologia")
# 129.1M params, eval mode, 12 relation labels over 13 entity types

Point it at a directory to load files you already have:

model, cfg = load_homologia("./homologia", device="cuda")

The loader rebuilds the architecture by pulling bowphs/GreBerta from the Hub, then loads this checkpoint over it with strict=True — every weight, encoder included, is replaced by the papyri-adapted one. First load therefore needs network access to the base repo. The transformers "newly initialized pooler" notice during that step is expected and harmless: the pooler is overwritten from the checkpoint a moment later, and strict=True would fail if anything were left uninitialised.

Candidate construction — which entity pairs to score, how to window a document longer than 512 tokens, how to fold per-window scores into one edge per pair — lives in oikonomia.relations.infer and oikonomia.relations.encode in the same package. Scoring raw pairs without that logic will not reproduce these numbers: the schema mask and the window merge are part of the model's decode.

Relation labels

HAS_AGE and HAS_OCCUPATION have no gold supervision. They were added to the label space from apposition rules after the evaluation below was run, so they are trained on silver only and are not scored anywhere on this card. They are predictable but unmeasured — treat them accordingly. The "gold edges" column is the count in the 115-document gold set (710 relations total) and is the single best predictor of which relations work.

Entity endpoints (13): AGE, COMMODITY, CURRENCY, DATE_REF, FRACTION, MONEY_AMOUNT, OCCUPATION, PERSON, PERSON_ROLE, QUANTITY, TAX_TERM, TRANSACTION, UNIT.

---

Training Details

Training Data

• Corpus: the Duke Databank of Documentary Papyri via papyri/idp.data, CC BY 3.0, pinned to revision d7a34f302d1e44e271256092c2b780733187b478. Not available as a Hub dataset; it is EpiDoc XML in a git repository.
• The papyri carry no relation markup upstream. All supervision was built for this project:
  • Silver: a deterministic lexicon + rules labeler over 48,891 documents.
  • Reference set: 115 documents — 710 relations, including 87 payment-direction edges adjudicated by verb class, not by grammatical case. Model-drafted and model-re-checked, not expert-validated.

Note on provenance. This reference set was drafted by one language model and re-checked span by span by a second, different one. It is mechanically constrained (offsets computed not typed, text byte-identical to the corpus, every numeral either labelled or explicitly skipped with a reason, every relation schema-legal) but no papyrologist or other domain expert has adjudicated it, and the maintainers do not read Ancient Greek. Scores below are therefore agreement with this reference, not accuracy against expert ground truth. Independent expert annotation is the top outstanding item.

Training Procedure

Silver pretraining of the span-pair head, then gold fine-tuning, over the DAPT-adapted encoder.

Training Hyperparameters

Head configuration: type_dim 64, feat_dim 16, dropout 0.2, schema-constrained decoding on, NO_RELATION class weight 1.0 (reweighting it was tested and made no difference).

Compute

Single NVIDIA A10 (24 GB) on Modal. Emissions were not tracked; a run of this size is on the order of GPU-hours, not GPU-days.

---

Evaluation

Testing Data, Factors & Metrics

5-fold cross-validation on the 115-document gold set. Two regimes, and the difference between them is the most important thing on this card:

• Oracle — scored with gold entity spans, isolating the relation model from entity errors. Every relation-extraction paper reports this; it flatters.
• End-to-end — scored on spans predicted by Grammateus, which is how you will actually use it.

Metric is micro-averaged F1 over typed, directed entity pairs.

Results — oracle entities

Per-relation F1 (oracle):

Results — end-to-end (quote these)

Run on Grammateus-predicted entities. The join between the two models is clean (0 documents unmatched).

† PARTY_OF oracle from the all-gold launch run's held-out CV — the honest generalization number for the shipped checkpoint — rather than the 0.652 in the table above, which is from the earlier paired-CV run.

The entity cascade costs ≈ 8 points on PARTY_OF (0.705 → 0.623). Note that recall, not precision, is what collapses end-to-end (P 0.771 / R 0.503): the model is still right about what it asserts, but the entity stage never hands it some of the pairs.

Four relations were not broken out individually in the end-to-end run and are reported only inside the 0.609 overall. That is a gap in the record.

One number deliberately not quoted

Scoring the shipped all-gold checkpoint on the gold documents gives PARTY_OF 0.993. That is train-on-test — those documents are in its training set — and it is reported here only because it confirms the save/load and constrained-decode inference path works end to end. It is not a generalization estimate. The 0.623 is itself mildly optimistic for the same reason, so true corpus-wide PARTY_OF is likely a shade under 0.62.

---

Bias, Risks, and Limitations

• Payment direction is weak, and this is data scarcity, not a bug. PAID_BY (0.145) and PAID_TO (0.300) sit far below the adjacency relations. The gold set holds only 87 direction edges — about 17 per held-out fold — and every model-side remedy tried (direction features, wide context, constrained decoding) measured neutral or slightly negative. We act on this: direction is deliberately absent from the published OIKONOMIA-DB. There are no paid_by/paid_to columns, because shipping them at F1 0.145 would be shipping noise as data.
• Rare relations collapse end-to-end. HAS_PRICE and CHARGED_UNDER fall to 0.0 with predicted entities, because the entity model rarely supplies their COMMODITY/TAX_TERM endpoints. The adjacency relations (HAS_UNIT, HAS_CURRENCY) are the robust ones. Performance is stratified by relation, not uniform — read the per-relation table before relying on any single edge type.
• Two relations are unmeasured. HAS_AGE and HAS_OCCUPATION have no gold supervision and no score.
• Errors are not uniform across documents. Formulaic contracts, which follow a fixed Greek template, are much easier than free-form letters and accounts. A corpus-wide average understates the first and overstates the second.
• Downstream social inference is not this model's claim. OIKONOMIA-DB combines PARTY_OF edges with rule-based gender attribution to study women as economic principals. Those are aggregate, error-bounded historical claims layered on top of this model, not assertions it makes about individuals.
• Corpus bias is inherited — surviving, published, digitized papyri, skewed toward the Arsinoite nome and dry sites.

Recommendations

Report end-to-end numbers, not oracle ones, if you build on this. Use the PARTY_OF/adjacency edges; treat direction as a hypothesis to verify. Prefer relative contrasts across periods, regions or document types to absolute counts, since extraction error is roughly common-mode across buckets.

---

Licence & lineage

Released under apache-2.0, inherited from the bowphs/GreBerta encoder (apache-2.0). The training corpus (DDbDP) is CC BY 3.0 and must be attributed. No ancestor carries a NonCommercial term — deliberately: the release is gated in code by a licence firewall that refuses any artifact of NonCommercial or unverified lineage (oikonomia.models.licensing; audit trail in MODEL_LICENSES.md).

Citation

@misc{oikonomia_homologia_2026,
  title  = {{OIKONOMIA-Homologia}: Relation Extraction for Greek Documentary Papyri},
  author = {Ainouche, Abderahmane},
  year   = {2026},
  url    = {https://huggingface.co/ainouche-abderahmane/homologia},
  note   = {Encoder bowphs/GreBerta; trained on the Duke Databank (DDbDP, CC BY 3.0)}
}

@inproceedings{riemenschneider-frank-2023-exploring,
  title     = {Exploring Large Language Models for Classical Philology},
  author    = {Riemenschneider, Frederick and Frank, Anette},
  booktitle = {Proceedings of ACL 2023},
  year      = {2023},
  url       = {https://arxiv.org/abs/2305.13698}
}

Duke Databank of Documentary Papyri (DDbDP), Duke Collaboratory for Classics Computing (DC3) and papyri.info, CC BY 3.0.

Model Card Contact

Issues and corrections: https://github.com/abderahmane-ai/oikonomia/issues