docs/prd/PRD_COGNATE_PAIRS_V2.md

PRD: Cognate Pairs Dataset v2 — Reconstruction from Verified Sources

Status: Complete Date: 2026-03-14 (implemented 2026-03-13 through 2026-03-14) Priority: P0 — Cognate pairs are used for validation testing; any hallucinated data invalidates model evaluation.

---

1. Problem Statement

Deep adversarial audit of the cognate pairs dataset (data/training/cognate_pairs/) has identified 6 critical bugs in the generation pipeline (scripts/assign_cognate_links.py and scripts/expand_cldf_full.py). Multiple bugs fabricate or mislabel cognate relationships, making the dataset unsuitable for validation. This PRD specifies the complete reconstruction of the cognate pairs from verified scholarly sources.

1.1 Current State

File	Rows	Size
cognate_pairs_inherited.tsv	18,257,301	1.18 GB
cognate_pairs_borrowing.tsv	116,757	8.0 MB
cognate_pairs_similarity.tsv	170,064	16.2 MB

Current 10-column schema:

Lang_A  Word_A  IPA_A  Lang_B  Word_B  IPA_B  Concept_ID  Relationship  Score  Source

1.2 Bugs Found

Bug 1: ABVD cognates.csv Never Read (CRITICAL)

Location: expand_cldf_full.py lines 200-266 Impact: All ABVD cognate data comes from forms.csv Cognacy column instead of the authoritative cognates.csv CognateTable (291,675 expert entries). The Doubt column (true/false) in cognates.csv is never consulted — disputed cognate assignments are treated as certain.

Bug 2: ABVD Multi-Set Cognacy Truncation

Location: expand_cldf_full.py line 257: cog_num = cognacy.split(",")[0].strip() Impact: Forms belonging to multiple cognate sets (e.g., "1,64") lose all secondary memberships. Estimated 37,587 cognate set memberships silently discarded.

Bug 3: WOLD Borrowing Pairs Fabricated (CRITICAL)

Location: assign_cognate_links.py lines 185-255 (_extract_borrowings_from_forms) Impact: The script reads forms.csv Borrowed column and pairs any two forms from different languages that share a concept AND have SCA similarity >= 0.3, labeling them "borrowing". This fabricates borrowing relationships. The authoritative borrowings.csv BorrowingTable (21,624 explicit donor-recipient events with Target_Form_ID, Source_Form_ID, Source_languoid, Source_certain fields) is COMPLETELY UNUSED.

Example: A Turkish word and an Arabic word for the same concept are paired as a "borrowing" even if no actual borrowing relationship exists between them.

Bug 4: Concept-Aligned Pairs Mislabeled as Inherited

Location: assign_cognate_links.py lines 154, 160 Impact: 488,900+ algorithmically-generated concept-aligned pairs (SCA similarity >= 0.5) are written to cognate_pairs_inherited.tsv with Relationship = cognate_inherited, making them indistinguishable from expert cognates to downstream consumers.

Bug 5: Sino-Tibetan Word Field is Concept String

Location: expand_cldf_full.py line 319 Impact: For all Sino-Tibetan entries, Word == Concept_ID (e.g., "above", "all") instead of the actual lexical form. The IPA column contains the real phonetic data, but Word is a meaningless concept gloss.

Bug 6: 50-Entry Hard Truncation (File-Sort Bias)

Location: assign_cognate_links.py line 142: members_sample = members[:50] Impact: For large families (Austronesian: hundreds of languages), only the first 50 entries in alphabetical ISO order are used. Languages late in the alphabet are systematically excluded.

---

2. Design: Reconstructed Pipeline

2.1 Iron Law Compliance

All data enters the dataset through code that reads from external scholarly sources. No hardcoded lexical or cognate data. No AI-generated entries.

Every extraction script:

• Reads from CLDF/TSV source files cloned to sources/
• Parses, transforms, writes to staging/cognate_pairs/
• Writes Source_Record_ID for full provenance traceability

2.2 Extended Schema (14 Columns)

Lang_A  Word_A  IPA_A  Lang_B  Word_B  IPA_B  Concept_ID  Relationship  Score  Source  Relation_Detail  Donor_Language  Confidence  Source_Record_ID

Column	Type	Description
Lang_A	str	ISO 639-3 code of language A
Word_A	str	Orthographic/transliteration form in language A
IPA_A	str	IPA transcription of Word_A
Lang_B	str	ISO 639-3 code of language B
Word_B	str	Orthographic/transliteration form in language B
IPA_B	str	IPA transcription of Word_B
Concept_ID	str	Concepticon gloss or equivalent concept identifier
Relationship	str	One of: expert_cognate, borrowing, concept_aligned, similarity_only
Score	float	SCA-based similarity score (0.0-1.0), rounded to 4 decimal places
Source	str	Source database identifier (e.g., abvd, wold, iecor, sinotibetan, acd)
Relation_Detail	str	Populated ONLY when source provides: inherited, borrowed, or -
Donor_Language	str	For borrowings only: source language from WOLD Source_languoid. - otherwise
Confidence	str	Source-provided certainty: certain/doubtful (ABVD), 1-5 (WOLD), - otherwise
Source_Record_ID	str	Traceable ID: ABVD cognateset ID, WOLD borrowing ID, IE-CoR cognateset ID, etc.

Design decisions:

• No mother-daughter / sister-sister typing — NO source provides this at the pair level.
• Relation_Detail, Donor_Language, Confidence are - when source doesn't provide them. We never fabricate metadata.
• Relationship is script-assigned based on extraction method. Relation_Detail is source-provided.

2.3 Source Inventory

Source	Repository	Cognate Data	Status
ABVD	sources/abvd/	cognates.csv CognateTable (291,675 entries, 19,356 sets)	Cloned, unused by current code
WOLD	sources/wold/	borrowings.csv BorrowingTable (21,624 events)	Cloned, unused by current code
Sino-Tibetan	sources/sinotibetan/	sinotibetan_dump.tsv (6,159 entries with COGID)	Cloned, partially used
IE-CoR	sources/iecor/	cognates.csv CognateTable (Indo-European cognates)	NOT CLONED — must clone
ACD	sources/acd/	Cached Austronesian data	Partially cached
Internal lexicons	data/training/lexicons/	SCA similarity scoring	N/A

2.4 Output Files

Same 3-file split, but with corrected labels:

• cognate_pairs_inherited.tsv — Expert cognates ONLY (ABVD CognateTable, IE-CoR CognateTable, Sino-Tibetan COGID, ACD)
• cognate_pairs_borrowing.tsv — Verified borrowings ONLY (WOLD BorrowingTable with explicit donor-recipient pairs)
• cognate_pairs_similarity.tsv — Concept-aligned pairs (algorithmically generated, clearly labeled)

---

3. Implementation Plan

Phase 1: Source Preparation

Step 1.1 — Clone IE-CoR:

cd sources/
git clone https://github.com/lexibank/iecor.git

Step 1.2 — Verify all CLDF source files exist:

• sources/abvd/cldf/cognates.csv (291,675 rows)
• sources/abvd/cldf/forms.csv
• sources/abvd/cldf/languages.csv
• sources/wold/cldf/borrowings.csv (21,624 rows)
• sources/wold/cldf/forms.csv
• sources/wold/cldf/languages.csv
• sources/sinotibetan/sinotibetan_dump.tsv (6,159 rows)
• sources/iecor/cldf/cognates.csv (NEW)
• sources/iecor/cldf/forms.csv (NEW)

Phase 2: Extraction Scripts

Each script follows the same pattern:

1. Read source CLDF files
2. Parse and validate
3. Generate pairwise cognate entries
4. Write to staging/cognate_pairs/{source}_pairs.tsv with 14-column schema
5. Print statistics and provenance summary

Script 1: scripts/extract_abvd_cognates_v2.py

Source: sources/abvd/cldf/cognates.csv + forms.csv + languages.csv

Key fixes:

• Read cognates.csv directly (has Form_ID, Cognateset_ID, Doubt)
• Join to forms.csv on Form_ID for orthographic form and IPA
• Handle multi-set membership: one form can appear in multiple cognate sets
• Use Doubt column for Confidence: Doubt=false → certain, Doubt=true → doubtful
• Source_Record_ID = Cognateset_ID from cognates.csv
• Generate cross-language pairs within each cognate set
• SCA similarity score computed on the fly

Script 2: scripts/extract_wold_borrowings_v2.py

Source: sources/wold/cldf/borrowings.csv + forms.csv + languages.csv

Key fixes:

• Read borrowings.csv BorrowingTable for actual donor-recipient pairs
• Join Target_Form_ID and Source_Form_ID to forms.csv for word/IPA
• Extract Source_languoid as Donor_Language
• Extract Source_certain for Confidence
• Source_Record_ID = borrowing ID from borrowings.csv
• Each row produces exactly ONE pair (not fabricated from concept co-occurrence)

Script 3: scripts/extract_sinotibetan_cognates_v2.py

Source: sources/sinotibetan/sinotibetan_dump.tsv

Key fixes:

• Filter out rows where BORROWING column is non-empty
• Use IPA column for IPA_A/IPA_B (correct)
• Use CONCEPT column for Concept_ID
• Use actual IPA form as Word (not concept gloss) — or use - and note that Word is not available
• Source_Record_ID = st_{COGID}

Script 4: scripts/extract_iecor_cognates.py (NEW)

Source: sources/iecor/cldf/cognates.csv + forms.csv

Implementation:

• Read IE-CoR CognateTable (standard CLDF format)
• Join Form_ID to forms.csv for word/IPA/language
• Generate cross-language pairs within each cognate set
• Source_Record_ID = IE-CoR Cognateset_ID

Script 5: scripts/extract_acd_cognates.py (NEW)

Source: sources/acd/ cached data

Implementation:

• Parse ACD (Austronesian Comparative Dictionary) cached HTML/data
• Extract cognate sets with etymon-level grouping
• If ACD data is insufficiently structured, mark as P2 and skip

Script 6: scripts/rebuild_concept_aligned_pairs.py

Source: data/training/lexicons/*.tsv + data/training/family_map.json

Key fixes:

• Label as concept_aligned (NOT cognate_inherited)
• Random sampling instead of file-sort truncation for groups > 50
• Score threshold: >= 0.5 → concept_aligned, 0.3-0.49 → similarity_only
• Source_Record_ID = - (no source record, algorithmically generated)
• Confidence = - (not from an expert source)

Script 7: scripts/merge_cognate_pairs.py

Merges all staging files into final output:

• Deduplicates: if pair (A,B,concept) appears in both expert and concept-aligned, keep expert
• Priority: expert > borrowing > concept_aligned > similarity
• Writes 3 output files with 14-column schema
• Prints final statistics

Phase 3: Adversarial Audit Protocol

For EACH extraction script, deploy a two-team audit:

Team A (Extraction):

• Runs the script
• Produces staging output
• Reports entry counts and statistics

Team B (Adversarial Auditor):

• Samples 20 random output rows
• For each row, traces Source_Record_ID back to source CSV
• Verifies Form_ID, Cognateset_ID, Language_ID all exist in source
• Verifies IPA matches source data
• Checks for entries in output that have no source backing
• Checks for duplicate pairs, empty fields, malformed data
• VETO power: if any entry cannot be traced, the entire script output is rejected

End-to-end cross-validation (after merge):

• Sample 50 random pairs from each output file (150 total)
• Full provenance trace: output → staging → source CSV → published database
• Verify no concept-aligned pairs appear in cognate_pairs_inherited.tsv
• Run count statistics and compare to source totals

Phase 4: Deployment

1. Commit PRD + all scripts + staging data
2. Run full pipeline: extract → audit → merge → validate
3. Update docs/DATABASE_REFERENCE.md with new cognate pair statistics
4. Push to GitHub
5. Push to HuggingFace via scripts/push_to_hf.py

---

4. Acceptance Criteria

Criterion	Metric
Zero fabricated pairs	Every pair traceable to source record via Source_Record_ID
Zero mislabeled relationships	Expert cognates in inherited.tsv ONLY; concept-aligned in similarity.tsv
WOLD borrowings use BorrowingTable	All borrowing pairs from borrowings.csv, not forms.csv
ABVD doubt flags preserved	Confidence column reflects Doubt field
Multi-set membership preserved	Forms in multiple cognate sets generate pairs for ALL sets
Sino-Tibetan borrowings excluded	Zero entries with BORROWING flag in inherited output
IE-CoR coverage	Indo-European expert cognates present in inherited output
Adversarial audit passes	0/150 sample pairs fail provenance trace

---

5. Non-Goals

• Mother-daughter / sister-sister relationship typing: No source provides this at the pair level. We encode only what sources give us.
• Etymological chain reconstruction: No source provides intermediary language chains. Out of scope.
• Cross-source conflation: If ABVD and IE-CoR both provide a cognate set for the same forms, both are kept (deduplicated by pair identity, priority to more specific source).
• Replacing existing lexicon files: This PRD only covers cognate pairs. Lexicon TSVs are not modified.

---

6. Risks

Risk	Mitigation
IE-CoR schema differs from ABVD	Read actual CLDF headers; adapt column names
ACD data too unstructured	Mark as P2; skip if insufficient
WOLD borrowings.csv has broken Form_IDs	Join validation: log and skip unresolvable IDs
Sino-Tibetan Word column unusable	Use IPA as the primary identifier; Word = -
Large pair counts overwhelm merge	Stream-process with generators; don't hold all pairs in memory