gmikros/kathnlp-treebank

kathnlp Katharevousa Greek Treebank

A Universal-Dependencies-style reference treebank for Katharevousa Greek, the archaizing official register used in 20th-century Greek law, administration, and parliamentary discourse. The treebank covers 1,697 sentences from written parliamentary questions of the early Third Hellenic Republic (1976–1977) and is released alongside the kathnlp parsing pipeline.

• Paper: A Reproducible Universal Dependencies-Style Pipeline for Katharevousa Greek Parliamentary Text (arXiv:2605.22978)
• Code: https://github.com/gmikros/katharevousa-nlp-tooling
• Companion model: gmikros/kathnlp-xlmr (release candidate, planned)

Dataset summary

Item	Count
Frozen reference sentences	1,697
Training sentences	1,357
Held-out test sentences	340
Held-out test tokens	4,093

The split uses seed 42 and is identical to the one used in every benchmark row of the accompanying paper. Use it to reproduce the published numbers and to evaluate new models on a comparable surface.

Files

File	Purpose
train.conllu / test.conllu	Raw UD CoNLL-U format for traditional NLP tooling.
train.jsonl / test.jsonl	One sentence per line; structured tokens. Powers the HF Dataset Viewer.
snapshot_manifest.json	Provenance manifest of the merged batch and retry files.
split.json	The seed-42 train/test split (sentence IDs).
annotation_schema.yaml	Katharevousa-aware annotation schema (UD v2 + sidecar fields).
reports/*.json	Per-model benchmark reports (spaCy, Stanza, custom Stanza, mBERT, XLM-R, feature-based).

Schema

The CoNLL-U columns follow UD v2: ID FORM LEMMA UPOS XPOS FEATS HEAD DEPREL DEPS MISC. The JSONL rows expose:

{
  "sent_id": "1977_reconstructed-...",
  "text": "Full sentence string.",
  "annotator": "gpt-5.5",
  "source_sheet": "1977_reconstructed",
  "source_row": 12,
  "tokens": [
    {"id": "1", "form": "Ἡ", "lemma": "ὁ", "upos": "DET", "feats": "Case=Nom|...", "head": 2, "deprel": "det", ...}
  ]
}

Curation and provenance

• Source archive. Greek parliamentary written questions from 1976–1977, OCR-derived and reconstructed for word-internal hyphenation and line-break artifacts. The same archive underlies the companion DSH digital-humanities study.
• Annotation. Schema-constrained LLM-assisted annotation (GPT-5 family) with structured JSON output, automatic validation, retry queues, and deterministic snapshotting. 1,565 sentences come from the main batch path; 132 are retry replacements that passed validation after the original batch failed.
• Status. The annotations are automatically validated, not fully expert-adjudicated. A philologist adjudication round is in progress and will be released as a versioned update.

How to load

Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("gmikros/kathnlp-treebank")
print(ds)
# DatasetDict({
#     train: Dataset({ features: [...], num_rows: 1357 }),
#     test:  Dataset({ features: [...], num_rows: 340  }),
# })

Raw CoNLL-U

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="gmikros/kathnlp-treebank",
    filename="train.conllu",
    repo_type="dataset",
)
text = open(path, encoding="utf-8").read()

Benchmark protocol

All published numbers in the accompanying paper use the fixed seed-42 split and identical scoring code (src/kathnlp/evaluation/metrics.py in the GitHub repository). The per-model reports under reports/ capture UPOS accuracy, weighted DEPREL F1, UAS, and LAS for spaCy Greek, Stanza Greek, Stanza Ancient Greek PROIEL, a custom-trained Stanza, mBERT, the XLM-R release candidate, and a feature-based logistic baseline.

Limitations

• Small held-out split (340 sentences / 4,093 tokens).
• Genre limited to parliamentary questions; transfer to legal texts, decrees, or newspapers is untested.
• Automatically validated rather than expert-adjudicated annotations.
• Polytonic and monotonic orthographic variation is preserved as in the source archive.

License

Released under Creative Commons Attribution-ShareAlike 4.0 (cc-by-sa-4.0), in line with most Universal Dependencies treebanks. The source archival material is in the public domain (Greek parliamentary records).

Citation

@misc{mikrosfitsilis2026kathnlp,
  title         = {A Reproducible Universal Dependencies-Style Pipeline for
                   Katharevousa Greek Parliamentary Text},
  author        = {Mikros, George and Fitsilis, Fotios},
  year          = {2026},
  eprint        = {2605.22978},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.22978}
}