# 003 — Cognate Pairs v2 Rebuild

**Date**: 2026-03-13
**Scope**: Complete rebuild of all cognate pair files, fixing 6 critical pipeline bugs

---

## 1. Objective

The original cognate pair extraction pipeline (v1) had 6 critical bugs that produced incorrect or incomplete data. This entry documents the complete rebuild from scratch using corrected v2 extraction scripts.

## 2. Bugs Fixed

| # | Bug | Impact | Root Cause |
|---|-----|--------|------------|
| 1 | ABVD never read `cognates.csv` | 291K expert cognate judgements missing | Script read `forms.csv:Cognacy` column instead of the dedicated `cognates.csv` CognateTable |
| 2 | Multi-set cognacy truncation | 37K lost set memberships | `Cognacy` field like `"1,2"` was truncated to `"1"` instead of splitting |
| 3 | WOLD fabricated borrowing pairs | False borrowing relationships | Script generated pairs from concept co-occurrence instead of reading `borrowings.csv` donor-recipient table |
| 4 | Concept-aligned pairs mislabeled as inherited | Inflated inherited count | Pairs from concept alignment (no expert cognacy) were labeled `expert_cognate` |
| 5 | Sino-Tibetan Word field = Concept string | Wrong word forms | `Word_A`/`Word_B` contained concept labels, not actual word forms |
| 6 | Alphabetical 50-entry hard truncation | Large families truncated | Families with >50 entries in a concept set had entries past the 50th alphabetically discarded |

## 3. Scripts Used

| Script | Purpose |
|--------|---------|
| `scripts/extract_abvd_cognates_v2.py` | Read authoritative `cognates.csv` (291K entries), handle multi-set membership, include Doubt column |
| `scripts/extract_wold_borrowings_v2.py` | Read authoritative `borrowings.csv` (21K donor-recipient events), extract Target_Form_ID + Source_Form_ID |
| `scripts/extract_sinotibetan_cognates_v2.py` | Read `sinotibetan_dump.tsv` (6,159 entries with COGID), separate inherited vs borrowing by BORROWING column |
| `scripts/extract_iecor_cognates.py` | Process IE-CoR CLDF CognateTable |
| `scripts/extract_acd_cognates.py` | Process Austronesian Comparative Dictionary |
| `scripts/merge_cognate_pairs.py` | Deduplicate across sources with priority ordering, produce 3 output files |

## 4. Data Sources

Same 6 CLDF sources as 001, plus IE-CoR and ACD from 002.

## 5. Source Reputability

- **ABVD `cognates.csv`**: 291,000 expert cognate judgements made by field linguists over 20+ years. Each entry has a `Doubt` column (certain/doubtful). This is the gold standard for Austronesian cognacy.
- **WOLD `borrowings.csv`**: 21,000+ explicit donor→recipient borrowing events curated by 41 specialist authors. Each has a certainty score (1-5 scale).
- **IE-CoR**: Expert cognate sets for Indo-European, curated by computational historical linguists at MPI-EVA.
- All entries have `Source_Record_ID` for traceability back to original CLDF records.

## 6. Methodology

### 14-Column Output Schema
```
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
```

### Scoring
- **SCA-weighted Levenshtein**: Normalized edit distance on Sound Class Alphabet (List 2012) encodings
- Substitution cost: 0 for same class, 0.5 for similar classes (e.g., voiced↔voiceless stop), 1.0 for different classes
- Gap penalty: 0.5
- Score = 1.0 − (normalized_distance), range [0.0, 1.0], rounded to 4 decimal places

### Deduplication Priority
When the same (Lang_A, Lang_B, Concept_ID) tuple appears in multiple sources:
1. `expert_cognate` (priority 0) — keeps expert-labelled pair
2. `borrowing` (priority 1)
3. `concept_aligned` (priority 2)
4. `similarity_only` (priority 3) — lowest priority, only kept if no better evidence

Pair key is order-independent: `min(side_a, side_b) || max(side_a, side_b) || concept`

### ABVD v2 Extraction (Bug #1 + #2 fix)
- Reads `cognates.csv` (CLDF CognateTable), not `forms.csv:Cognacy`
- Splits multi-set membership: `"1,2"` → entries in both cognateset 1 and 2
- Preserves `Doubt` column as `Confidence` field

### WOLD v2 Extraction (Bug #3 fix)
- Reads `borrowings.csv` directly: each row = one donor→recipient event
- Extracts `Target_Form_ID` and `Source_Form_ID`, resolves to language/word/IPA
- No fabricated pairs — only explicit borrowing relationships

## 7. Tests Performed

- All 406 existing tests pass after rebuild
- `test_full_pipeline.py`: End-to-end Ugaritic-Hebrew cognate detection verified
- `test_training_data.py`: Schema validation on all 3 output files
- Row count verification against source repos:
  - ABVD: `wc -l cognates.csv` matches expected extraction count
  - WOLD: `wc -l borrowings.csv` matches expected extraction count
- Deduplication audit: Verified no duplicate (Lang_A, Lang_B, Concept_ID) tuples in output

## 8. Cross-Referencing

- 20 random inherited pairs traced back to ABVD `cognates.csv` by `Source_Record_ID`
- 10 random borrowing pairs traced back to WOLD `borrowings.csv` by `Source_Record_ID`
- Verified WOLD borrowing pairs have correct `Donor_Language` by cross-referencing with source
- Verified Sino-Tibetan Word fields now contain actual word forms (not concept labels)

## 9. Output Summary

| File | Rows | Size | Change from v1 |
|------|------|------|-----------------|
| `cognate_pairs_inherited.tsv` | 22.9M | 2.2 GB | +291K expert cognates recovered |
| `cognate_pairs_borrowing.tsv` | 17K | 1.9 MB | Completely rebuilt (v1 was fabricated) |
| `cognate_pairs_similarity.tsv` | 465K | 49.9 MB | Relabeled (was falsely `inherited` in v1) |

## 10. PRD Reference

Full specification: `docs/prd/PRD_COGNATE_PAIRS_V2.md`