tadad/diorisis-ancient-greek

Diorisis Ancient Greek Corpus

A Hugging Face conversion of Alessandro Vatri and Barbara McGillivray's Diorisis Ancient Greek Corpus for the BigLAM community. Diorisis contains 820 literary texts from Homer through the fifth century CE, with automatic lemma, part-of-speech, and morphological annotations.

The conversion combines the original XML headers with the JSON corpus and a checksum-pinned snapshot of the author's public per-file corrections. It retains Beta Code and adds Unicode Greek.

Structure

Configuration	Unit	Rows	Contents
sentences	Sentence	538,001	Ordered annotated tokens and work metadata
works	Work	820	Bibliographic metadata, counts, and reconstructed text

sentences is the default configuration. Each row contains normalized Beta Code and Unicode text plus a tokens list. Token records preserve the source form, lemma identifiers and entries, parts of speech, possible morphological analyses, TreeTagger and disambiguation values, quote and lacuna flags, poetry line-break markers, and unrecognized source fields. lemmas retains every candidate when a correction contains more than one; the singular convenience fields are null rather than choosing a candidate.

works contains author, title, TLG identifiers, approximate composition year, genre, subgenre, source link, license statement, corpus version, counts, reconstructed text, and the source header. work_id includes the title because some corpus files share a TLG author/work pair.

from datasets import load_dataset

sentences = load_dataset("tadad/diorisis-ancient-greek", "sentences")
works = load_dataset("tadad/diorisis-ancient-greek", "works")

The source supplies no evaluation split. All rows use train; downstream projects should split by work_id, author, or composition period to prevent sentence-level leakage.

Source versions and conversion

The original 2018 XML release supplies TEI headers and the initial annotations. The fixed Figshare JSON release supplies corpus version 1.51. The author's public updater supplies later per-file corrections; source/update_manifest.json pins the version and checksum of each applied payload. source/diorisis_updates.zip stores those payloads so the snapshot remains rebuildable if the update service changes.

The snapshot contains 497 works at corpus version 1.6 and 323 at version 1.51. Together they contain 10,203,006 word tokens and 1,337,457 punctuation tokens.

The builder verifies both Figshare downloads, matches all JSON works to their XML headers, converts Beta Code to NFC Unicode, preserves token order and annotation alternatives, and writes typed Parquet. reports/build-report.json records corpus statistics and output checksums.

Limitations

• Lemma, part-of-speech, morphology, and sentence boundaries are automatic annotations rather than reviewed ground truth.
• Composition dates are approximate and often represent conventional scholarly dates.
• The collection reflects the coverage and editorial choices of Perseus, Bibliotheca Augustana, and Mikros Apoplous.
• Reconstructed text_beta and text_unicode use normalized whitespace because the tokenized source does not preserve original spacing.
• The source contains unresolved and alternative morphological analyses. These remain in the analyses list.

Rebuild

uv sync
uv run python -m conversion.build_dataset
uv run python -m conversion.validate_dataset
uv run pytest

Pass --refresh-updates only when intentionally creating a new correction snapshot. Normal builds use the checked-in update manifest.

License and citation

The Figshare records declare CC BY 4.0. The corpus manifest says CC-BY-SA, and the XML work headers identify CC BY-SA 3.0 US. This dataset preserves the per-work statements and uses license: other so that the metadata does not erase that discrepancy. Users should follow the most restrictive applicable source terms.

@article{vatri_mcgillivray_2018_diorisis,
  author  = {Vatri, Alessandro and McGillivray, Barbara},
  title   = {The Diorisis Ancient Greek Corpus},
  journal = {Research Data Journal for the Humanities and Social Sciences},
  volume  = {3},
  number  = {1},
  pages   = {55--65},
  year    = {2018},
  doi     = {10.1163/24523666-01000013}
}

This conversion responds to BigLAM issue #65.