Hugging Face's logo

Datasets:
Nacryos
/
ancient-scripts-datasets

Tasks:
Text Classification
Token Classification
Languages:
Multiple languages
Size:
10M<n<100M
ArXiv:
Tags:
linguistics
historical-linguistics
cognate-detection
phylogenetics
ancient-languages
IPA
License:
Dataset card Data Studio Files
xet
Community
1
ancient-scripts-datasets / docs /prd /PRD_LEXICON_EXPANSION.md
Alvin
Add complete dataset: all sources, metadata, scripts, docs, and phylo enrichment
26786e3
preview code
|
raw
history blame contribute delete
7.91 kB

PRD: Comprehensive Ancient Language Lexicon Expansion

IRON RULE: NO DATA GENERATION — EXTRACTION ONLY

This rule supersedes ALL other goals including target lemma counts.

  1. NEVER generate, fabricate, or hallucinate any lexical data. Every entry must trace to a verifiable external source.
  2. NEVER write TSV rows directly. Only write CODE (Python scripts) that fetches from external sources.
  3. Every script must use requests.get(), urllib, BeautifulSoup, or similar HTTP extraction. Never f.write("word\tIPA\t...") with hardcoded linguistic content.
  4. If a source is unreachable or returns fewer entries than expected: report the actual count. NEVER pad with invented entries.
  5. Target lemma counts in the table below are ASPIRATIONAL ESTIMATES, not quotas. If Wiktionary has 200 Avestan lemmas (not 400), we get 200. The count column exists only to prioritize effort.
  6. If in doubt whether data is real: skip it, log it, flag it for manual review.
  7. Adversarial auditors have VETO POWER. If an auditor flags an entry as unverifiable, it is REMOVED.

Context

All 18 ancient language lexicons now PASS adversarial audit, but entry counts are low relative to what's available online. Research agents identified major untapped sources with 2-10x more data than we currently have. Root causes: artificial fetch limits, missing dedicated-site parsers, incomplete Wiktionary category pagination, and hardcoded-only extraction for some languages.

Current vs Available (researched counts):

Language ISO Current Wiktionary Available Dedicated Source Aspirational (not quota)
PIE ine 841 1,704 lemmas Palaeolexicon 1,000+ 1,500+
Ugaritic uga 344 761 lemmas — 600+
Old Persian peo 244 524 lemmas — 400+
Avestan ave 157 400 lemmas avesta.org 3,000+ 500+
Proto-Dravidian dra 171 367 lemmas DEDR 5,569 groups 400+
Hittite hit 251 251 lemmas AssyrianLanguages.org 300+
Proto-Semitic sem 139 248 lemmas — 200+
Proto-Kartvelian ccs 100 250 lemmas — 200+
Phoenician phn 181 166 lemmas FREELANG 200+
Elamite elx 301 ~300 — 301 (maxed)
Phrygian xpg 73 62 lemmas Obrador-Cursach 100+
Urartian xur 44 ~50 Oracc eCUT 742 texts 150+
Lycian xlc 75 105 lemmas eDiAna 3,757 total 150+
Lydian xld 58 65 lemmas eDiAna 100+
Carian xcr 54 51 lemmas eDiAna 80+
Lemnian xle 33 ~30 — 33 (maxed)
Messapic cms 46 ~40 — 50+
Rhaetic xrr 37 ~30 TIR 389 inscriptions 60+

Phase 1: Fix Existing Parsers & Remove Limits

1a. Remove artificial fetch limits

Files to modify:

  • scripts/extract_ave_peo_xpg.py: Remove max_fetch=50 (ave), max_fetch=30 (xpg)
  • scripts/parsers/parse_lrc.py: Remove max_pages=30 limit
  • scripts/parsers/parse_dedr.py: Remove max_pages=10 limit

1b. Fix Wiktionary category pagination in parse_wiktionary.py

Currently fetches only first page of category members. Fix: use cmcontinue token to paginate through all results. This alone should double PIE, Ugaritic, Old Persian, Avestan, Proto-Semitic, Proto-Kartvelian entries.

1c. Fix phn.tsv gloss quality

The v1 agent found 166/181 glosses (91.7%) vs v2's 123/181 (68%). Check current TSV state and apply whichever result is better.

Phase 2: Wiktionary Deep Scrape (7 languages, parallel)

For each language, create a dedicated expansion script that:

  1. Paginates through ALL Wiktionary category members (not just first page)
  2. Fetches each word page for gloss extraction
  3. Handles language-specific properties (abjad scripts, cuneiform, reconstructed forms)
  4. Deduplicates against existing TSV entries
  5. Applies transliteration→IPA via existing transliteration_maps.py
  6. Computes SCA via existing pipeline

Language-specific context:

Language Script/Encoding Notes Category
PIE Reconstructed forms with *asterisk prefix, laryngeals (h₁h₂h₃) Category:Proto-Indo-European_lemmas
Ugaritic Cuneiform transliteration (Bordreuil/Pardee conventions), 30-letter alphabet Category:Ugaritic_lemmas
Old Persian Cuneiform syllabary (Kent notation), word dividers Category:Old_Persian_lemmas
Avestan Avestan script entries may be under script or transliteration Category:Avestan_lemmas
Proto-Dravidian DEDR notation with *asterisk, retroflex series (ṭ ḍ ṇ ḷ) Category:Proto-Dravidian_lemmas
Proto-Semitic Reconstructed roots with *asterisk, triconsonantal structure Category:Proto-Semitic_lemmas
Proto-Kartvelian Klimov notation, ejective series (p' t' k' q') Category:Proto-Kartvelian_lemmas

Phase 3: Dedicated Source Scraping (5 languages, parallel)

3a. Avestan — avesta.org dictionary

  • Source: avesta.org/dictionary/ — comprehensive Avestan-English dictionary (3,000+ headwords)
  • Parser: Extract headwords + glosses from HTML dictionary pages

3b. Urartian — Oracc eCUT

  • Source: Oracc Electronic Corpus of Urartian Texts (JSON API)
  • Try all 3 JSON strategies: glossary, index/lem, members/cdl
  • Expected: 150+ unique lemmas from 742 texts

3c. Anatolian trio (Lycian/Lydian/Carian) — Palaeolexicon

  • Source: palaeolexicon.com — searchable ancient language database
  • Parser: Extract word lists per language with glosses

3d. Phrygian — Wiktionary + deep scrape

  • Current: 73 entries from 62 Wiktionary lemmas + hardcoded supplement
  • Expand: paginate fully, also check Phrygian appendix pages

3e. Rhaetic — TIR (Thesaurus Inscriptionum Raeticarum)

  • Source: univie.ac.at/raetica/ — Semantic MediaWiki with 389 inscriptions
  • Parser: Extract word forms from inscription pages

Phase 4: Integration & Verification

4a. Merge all new data

  • Deduplicate by Word column (case-sensitive)
  • Keep existing entries unchanged; only ADD new ones
  • Verify IPA and SCA for all new entries

4b. Re-run adversarial audits

4c. Update metadata

4d. Commit and push

Files to Modify

Action File
MODIFY scripts/extract_ave_peo_xpg.py (remove fetch limits)
MODIFY scripts/parsers/parse_wiktionary.py (add pagination)
MODIFY scripts/parsers/parse_lrc.py (remove page limit)
MODIFY scripts/parsers/parse_dedr.py (remove page limit)
CREATE scripts/expand_wiktionary_{iso}.py (7 scripts, one per Phase 2 language)
CREATE scripts/scrape_avesta.py
CREATE scripts/scrape_oracc_urartian.py
CREATE scripts/scrape_palaeolexicon.py
CREATE scripts/expand_xpg.py
CREATE scripts/scrape_tir_rhaetic.py
MODIFY data/training/lexicons/*.tsv (18 files, add new entries)
MODIFY docs/ADVERSARIAL_AUDIT_*.md (18 files, update counts)
MODIFY data/training/metadata/languages.tsv

Existing Code to Reuse

Utility File Purpose
transliterate(text, iso) scripts/transliteration_maps.py All IPA conversion
ipa_to_sound_class() cognate_pipeline/ SCA generation
ALL_MAPS scripts/transliteration_maps.py Per-language transliteration dicts
language_configs scripts/language_configs.py ISO codes, source URLs
Existing parsers scripts/parsers/*.py Template for new parsers

Verification

  1. Each extraction agent produces a JSONL audit trail
  2. Each adversarial auditor samples 20 random entries and verifies against source
  3. All 18 audits must PASS
  4. Entry counts must increase (never decrease) for all languages
  5. No Word==IPA regression
  6. git diff --stat confirms only expected files changed
  7. Commit and push to GitHub