Alvin
Add inscription text + Phonetic Prior validation datasets
b807cee

Ancient Inscription Text Dataset

Running text (word sequences preserving original word order) from academically peer-reviewed, CC-BY-SA-compatible sources. Designed for phonotactic sequence analysis in Phonetic Prior models.

Languages

ISO Language Source Entries Date Range Genre
grc Ancient Greek UD_Ancient_Greek-PTNK (Septuagint) 1,576 4th-5th c. CE religious
lat Latin UD_Latin-LLCT + UD_Latin-CIRCSE 10,687 774-897 CE / 1st c. BCE-2nd c. CE legal + literary
san Sanskrit UD_Sanskrit-Vedic 27,182 c. 1500-500 BCE religious
ang Old English UD_Old_English-Cairo 20 c. 700-1100 CE mixed
osc Oscan CEIPoM v1.3 (Zenodo) 770 5th-1st c. BCE inscription
xum Umbrian CEIPoM v1.3 (Zenodo) 761 7th-1st c. BCE inscription

Total: ~41,000 entries across 6 languages.

TSV Schema

Column Type Description
Inscription_ID str Unique identifier (source + sentence ID)
Text str Original text (space-separated words)
IPA str IPA transcription (space-separated words)
SCA str Sound class encoding (space-separated words)
Source str Source database identifier
Date_Approx str Approximate date of the text
Genre str Text genre (legal, literary, religious, inscription)
IPA_Source str IPA generation method

Sources & Licenses

All sources are CC BY-SA 4.0 or more permissive:

IPA Transliteration

Language Method Reference
grc Greek Unicode -> strip diacritics -> transliteration map Allen (1987) Vox Graeca
lat Latin script -> transliteration map (Classical values) Allen (1978) Vox Latina
san IAST romanization -> transliteration map Whitney (1889) Sanskrit Grammar
ang Old English script -> transliteration map Campbell (1959) OE Grammar
osc CEIPoM Standard_aligned -> transliteration map Buck (1904); Pitts (2022)
xum CEIPoM Standard_aligned -> transliteration map Buck (1904); Pitts (2022)

SCA encoding uses the system from List (2012), implemented in cognate_pipeline/src/cognate_pipeline/normalise/sound_class.py.

Limitations

  1. 6 of 14 validation languages only: Gothic, OCS, Old Norse, Old Irish, Avestan, OHG, Old Turkic, Hebrew lack CC-BY-SA inscription corpora.
  2. Late Latin bias: LLCT is AD 774-897 Tuscan legal Latin, not Classical. CIRCSE adds Classical literary texts but is smaller.
  3. Old English is minimal: Only 20 sentences from Cairo treebank. OEDT has no released data yet.
  4. Oscan: Greek-script inscriptions excluded: ~240 Oscan inscriptions in Greek alphabet were filtered out (need separate transliteration path).
  5. IPA is approximate: Transliteration maps apply default/canonical pronunciations. Allophonic variation, sandhi, and contextual rules are not modeled.

Build Instructions

# Step 1: Download sources
python scripts/ingest_inscriptions.py

# Step 2: Build TSVs
python scripts/build_inscriptions.py

Citation

If using this dataset, cite the original sources listed above plus:

  • Pitts, R.J. (2022). The Corpus of the Epigraphy of the Italian Peninsula in the 1st Millennium BCE. Journal of Open Humanities Data, 8(7).
  • List, J.-M. (2012). SCA: A Method for Automatic Sound Correspondence Analysis.