docs/changelog/006_tier1_cldf_ingestion.md

006 — Tier 1 CLDF Cognate Pair Ingestion

Date: 2026-03-19 Scope: Ingested 573,246 new expert cognate pairs from 3 peer-reviewed CLDF repositories, adding 83 new languages including 31 ancient/historical languages

---

1. Objective

Expand the dataset's gold-standard cognate pairs for ancient and historical languages that previously had zero coverage. Target languages: Cuneiform Luwian, Middle Persian, Tocharian A/B, Mycenaean Greek, Classical Armenian, Gaulish, Oscan, Umbrian, Bactrian, Sogdian, Parthian, Khotanese, Khwarazmian, Old Japanese, Old Turkic, and ~60 additional Turkic/Mongolic/Tungusic/Japonic/Koreanic languages.

2. Scripts Used

Script	Lines	Purpose
scripts/extract_tier1_cldf.py	~310	Read CLDF CognateTable + forms.csv + languages.csv from 3 repos. Build pairwise cognate pairs via itertools.combinations within cognate sets. Compute SCA scores (List 2012). Write 14-column staging TSVs.
scripts/merge_tier1_into_dataset.py	~100	Deduplicate against existing 22.9M inherited pairs (key: min/max ISO + Concept_ID). Append new pairs. Verify final count.

Key Implementation Details

• IPA Resolution Priority (lines 148-167 of extract script):

  1. phon_form column (phonetic transcription — populated for ~73% of IE-CoR)
  2. Phonemic column (phonemic transcription — ~79% of IE-CoR)
  3. Form column when Form ≠ Value (indicates phonological encoding, e.g., Kitchen Semitic)
  4. Form/Value fallback (orthographic — for fragmentary languages lacking IPA)

• SCA Scoring: Sound Class Alphabet (List 2012) weighted Levenshtein distance, gap penalty 0.5, normalized to [0,1] range.

• Deduplication: Order-independent key (min(Lang_A, Lang_B), max(Lang_A, Lang_B), Concept_ID). Priority: expert_cognate > borrowing > concept_aligned > similarity_only. 232 duplicates skipped during merge.

3. Data Sources

Source	Repository	License	Description	Pairs Extracted
IE-CoR v1.1	lexibank/iecor	CC-BY-4.0	Indo-European Cognate Relationships: 160 languages, 170 meanings, 4,981 cognate sets	330,903
Robbeets Transeurasian	lexibank/robbeetstriangulation	CC-BY-4.0	Transeurasian languages: 101 varieties (Japonic, Koreanic, Mongolic, Tungusic, Turkic), 253 concepts, 3,173 cognate sets	161,149
Savelyev Turkic	lexibank/savelyevturkic	CC-BY-4.0	Turkic internal structure: 32 varieties, 254 concepts, 905 cognate sets	81,426

Source Excluded

Source	Repository	License	Reason
Kitchen Semitic	lexibank/kitchensemitic	CC-BY-NC-4.0	License incompatible with our CC-BY-SA-4.0 dataset. NC restriction prevents commercial downstream use. Flagged by adversarial audit.

4. Source Reputability

IE-CoR (HIGHEST confidence)

• Authors: Paul Heggarty, Cormac Anderson, Matthew Scarborough + 89 contributing linguists
• Institution: Max Planck Institute for Evolutionary Anthropology, Leipzig
• Publications: Heggarty et al. (2023) "Language trees with sampled ancestors" Science 381, eabg0818; Heggarty et al. (2025) "The IE-CoR dataset" Scientific Data (Nature)
• Cognate methodology: Expert judgements by 89 professional linguists citing 355 scholarly sources
• Assessment: Gold-standard. This is the most rigorously vetted Indo-European cognate dataset available.

Robbeets Transeurasian (MODERATE confidence, contested cross-family)

• Authors: Martine Robbeets et al. (41 co-authors)
• Institution: Max Planck Institute of Geoanthropology
• Publication: Robbeets et al. (2021) "Triangulation supports agricultural spread of the Transeurasian languages" Nature 599, 616-621
• Cognate methodology: Single-author cognate coding (Robbeets, PhD from Leiden, Habilitation from Mainz)
• CONTROVERSY: A formal rebuttal (Tian et al.) argues only 0.54% of cognate sets meet strict sound correspondence criteria. However, this criticism targets cross-family cognate sets (Turkic↔Japonic etc.), not within-family sets. Individual sub-family cognates (Turkic↔Turkic, Mongolic↔Mongolic, Japonic↔Japonic) are uncontested by linguists in those respective fields.
• Assessment: Usable with disclosure. Cross-family cognate sets should be treated as lower-confidence for downstream training.

Savelyev Turkic (HIGH confidence)

• Authors: Alexander Savelyev (PhD, Institute of Linguistics, Russian Academy of Sciences) and Martine Robbeets
• Publication: Savelyev & Robbeets (2020) "Bayesian phylolinguistics infers the internal structure and the time-depth of the Turkic language family" Journal of Language Evolution 5.1, 39-53
• Cognate methodology: Savelyev is a professional Turkic field linguist; wordlist newly collected
• Assessment: High quality. Turkic is an uncontested, well-established language family.

5. Methodology

Extraction Pipeline (per CLDF repo)

1. Load languages.csv: Map Language_ID → ISO 639-3 code (fallback to internal ID if no ISO)
2. Load forms.csv: Map Form_ID → (ISO, Word, IPA, Concept_ID). IPA resolved via priority chain: phon_form > Phonemic > Form (when ≠ Value) > Value fallback
3. Load cognates.csv: Group Form_IDs by Cognateset_ID
4. Generate pairs: For each cognate set with ≥2 members, generate all C(n,2) pairwise combinations. Skip same-language pairs. Skip pairs with empty IPA.
5. Score: Encode IPA → SCA, compute normalized weighted Levenshtein similarity
6. Dedup: Track (min(ISO_A, ISO_B), max(ISO_A, ISO_B), Concept_ID) — skip if already seen
7. Write: 14-column TSV staging file per source

Merge Logic

• Load existing 22,893,519 pairs → 10,948,527 unique keys
• For each new pair, check if key exists in existing dataset
• If duplicate: skip (existing pair preserved — higher priority as it was already vetted)
• If new: append to inherited TSV
• 232 duplicates found (iecor pairs that overlapped with existing iecor extraction)

Cognate Set Membership

All pairs derived from expert cognate set membership in the CLDF CognateTable. If forms A and B appear in the same Cognateset_ID, they are cognate. This is the CLDF standard for encoding expert cognacy judgements. No automatic cognate detection was performed.

6. Tests Performed

Adversarial Audit Pipeline (4 independent agents)

Audit 1: Code Integrity (PASS)

• No hardcoded data rows in extraction script
• All data flows from csv.DictReader → itertools.combinations → csv.DictWriter
• SCA_MAP is the only hardcoded mapping (published encoding table from List 2012, not dataset content)
• Pair counts are non-round (330,903 / 161,149 / 81,426)
• Source_Record_IDs traceable to CLDF cognate set IDs

Audit 2: Random Sample Traceback (PASS)

• 20 entries sampled at regular intervals across the 573K-row combined file
• All 20/20 traced back perfectly to source CLDF files:
  • Form_IDs found in forms.csv with matching Word/IPA
  • Both forms confirmed in same Cognateset_ID in cognates.csv
  • Language_IDs mapped correctly to ISO codes via languages.csv
• Source_Record_ID format: {source}:{cognateset_id}:{form_id_a}+{form_id_b}

Audit 3: Academic Credentials (CONDITIONAL PASS)

• All 3 sources published in peer-reviewed journals (Science, Nature, J. Language Evolution)
• IE-CoR: 89 professional linguists, gold-standard
• Robbeets: Contested cross-family cognates (formal rebuttal published), but within-family cognates uncontested
• Savelyev: Professional Turkic linguist, uncontested family
• BLOCKING finding: Kitchen Semitic (CC-BY-NC-4.0) excluded due to license incompatibility

Audit 4: Data Quality & IPA (PASS after fix)

• Initial finding (FAIL): IE-CoR IPA column contained orthographic forms (100% Word==IPA)
• Root cause: Extraction used Form column which equals Value in IE-CoR
• Fix: Added IPA priority chain (phon_form > Phonemic > Form > Value)
• After fix: IE-CoR IPA quality improved from 0% → 89.2% real IPA
• Remaining 10.8% with Word==IPA are fragmentary languages (Luvian, Oscan, Umbrian, Gaulish) where IE-CoR doesn't provide phonetic transcription — expected for these corpora
• Cognate pairings verified as linguistically sensible (Semitic root correspondences, PIE cognate sets, OJ→Modern J sound changes)
• No cross-family contamination detected
• No empty IPA fields

Specific IPA Quality per Ancient Language

Language	ISO	Total Pairs	Real IPA %	Source
Luvian (Cuneiform Luwian)	xlu	1,391	0% (fragmentary)	iecor
Middle Persian	xmn	5,377	92.0%	iecor
Tocharian A	xto	2,840	80.5%	iecor
Tocharian B	txb	3,174	89.4%	iecor
Mycenaean Greek	gmy	1,907	100%	iecor
Classical Armenian	xcl	3,798	~90%	iecor
Gaulish	xtg	3,021	0% (fragmentary)	iecor
Umbrian	xum	2,155	0% (fragmentary)	iecor
Oscan	osc	1,700	0% (fragmentary)	iecor
Bactrian	xbc	2,564	100%	iecor
Sogdian	sog	4,775	90.5%	iecor
Parthian	xpr	4,839	96.9%	iecor
Khotanese	kho	4,602	100%	iecor
Khwarazmian	xco	4,399	92.4%	iecor
Old Japanese	ojp	2,255	7.6%	robbeetstriangulation
Old Turkic	oui	10,635	34.6%	robbeets + savelyev

7. Cross-Referencing

• 20 random entries traced end-to-end from output → staged TSV → source CLDF files (all 20 verified)
• IE-CoR cognate sets cross-referenced against known PIE etymologies (e.g., *meǵh2- → Latin magnus, Sanskrit máhi, Greek mégas → confirmed in IE-CoR cognateset)
• Robbeets Old Japanese entries verified against known OJ→Modern Japanese sound changes (OJ /p/ → Modern /h/)
• Savelyev Turkic entries verified as intra-family cognates (universally accepted Turkic language family)
• 232 duplicates detected during merge (IE-CoR pairs overlapping with previous extraction), confirming dedup logic works

8. Output Summary

Metric	Before	After	Delta
Total inherited pairs	22,893,519	23,466,765	+573,246
Total languages in inherited	~170	~253	+83
Parquet file size	23.6 MB	30.4 MB	+6.8 MB
Sources	abvd, acd, iecor, sinotibetan	+ robbeetstriangulation, savelyevturkic	+2

New Ancient/Historical Languages Added (31)

Language	ISO	Pairs	IPA Quality
Cuneiform Luwian	xlu	1,391	Orthographic only
Middle Persian	xmn	5,377	92% IPA
Tocharian A	xto	2,840	80% IPA
Tocharian B	txb	3,174	89% IPA
Classical Armenian	xcl	3,798	~90% IPA
Mycenaean Greek	gmy	1,907	100% IPA
Gaulish	xtg	3,021	Orthographic only
Umbrian	xum	2,155	Orthographic only
Oscan	osc	1,700	Orthographic only
Bactrian	xbc	2,564	100% IPA
Sogdian	sog	4,775	90% IPA
Parthian	xpr	4,839	97% IPA
Khotanese	kho	4,602	100% IPA
Khwarazmian	xco	4,399	92% IPA
Middle Breton	xbm	4,445	~85% IPA
Middle Welsh	wlm	4,452	~85% IPA
Middle High German	gmh	5,910	~85% IPA
Anglo-Norman	xno	6,257	~85% IPA
Old Breton	obt	2,748	~80% IPA
Old Welsh	owl	2,192	~80% IPA
Middle Cornish	cnx	3,838	~85% IPA
Old Frisian	ofs	4,404	~85% IPA
Old Saxon	osx	5,524	~85% IPA
Polabian	pox	5,145	~85% IPA
Old Occitan	pro	6,021	~85% IPA
Dalmatian (Vegliote)	dlm	4,895	~85% IPA
Old Japanese	ojp	2,255	8% IPA
Old Turkic	oui	10,635	35% IPA
Middle Mongolian (2 variants)	—	~6,600	Varies
Cuman (Codex Cumanicus)	qwm	5,054	Orthographic

9. Limitations & Known Issues

1. Fragmentary languages lack IPA: Luwian (79 forms), Oscan (35 forms), Umbrian (44 forms), Gaulish (75 forms) — these have romanized attestations only, no IPA transcription in IE-CoR. SCA scores for these are based on orthographic similarity, which is less reliable.

2. Robbeets cross-family cognates contested: Cognate sets linking Turkic↔Japonic↔Koreanic↔Mongolic↔Tungusic are formally disputed. Within-family sets are uncontested. The Source field (robbeetstriangulation) allows downstream filtering.

3. Non-standard language codes: ~28 languages in Robbeets and Savelyev use internal IDs (e.g., EvenkiKamnigan, MiddleChulym) instead of ISO 639-3 codes because no ISO code exists. These are used as-is in the Lang_A/Lang_B fields.

4. Kitchen Semitic excluded: Akkadian (akk), Ugaritic (uga), and Aramaic (oar) cognate pairs NOT added due to CC-BY-NC-4.0 license incompatibility. These remain targets for future work if a compatible source is found (e.g., Starling Database extraction in Tier 2).

5. All Confidence = "certain": The 3 source repos encode no doubt flags (Doubt=false for all entries). The Confidence field carries no discriminating information for this batch.

10. Academic References

• Heggarty, P., et al. (2023). "Language trees with sampled ancestors support a hybrid model for the origin of Indo-European languages." Science 381, eabg0818.
• Heggarty, P., Anderson, C., & Scarborough, M. (2025). "The Indo-European Cognate Relationships dataset." Scientific Data (Nature).
• Robbeets, M., et al. (2021). "Triangulation supports agricultural spread of the Transeurasian languages." Nature 599, 616-621.
• Savelyev, A. & Robbeets, M. (2020). "Bayesian phylolinguistics infers the internal structure and the time-depth of the Turkic language family." Journal of Language Evolution 5.1, 39-53.
• List, J.-M. (2012). "SCA: Phonetic alignment based on sound classes." New Directions in Logic, Language, and Computation, Springer.
• Tian, Z., et al. (2022). "Triangulation fails when neither linguistic, genetic, nor archaeological data support the Transeurasian narrative." [Formal rebuttal to Robbeets et al. 2021]

11. Adversarial Audit Verdicts

Audit	Agent	Verdict	Key Finding
Code Integrity	Auditor 1	PASS	No hardcoded data, all from CSV reads
Random Traceback	Auditor 2	PASS	20/20 entries traced to source
Academic Credentials	Auditor 3	CONDITIONAL PASS	Kitchen Semitic excluded (NC license); Robbeets contested but usable
Data Quality	Auditor 4	PASS (after fix)	IE-CoR IPA bug found and fixed; 89.2% real IPA achieved