Fix Linear B Wiktionary tr= extraction bug causing 5 wrong IPA values

Root cause: re.search(r"|tr=([^|}]+)") matched the first tr= in the
entire wikitext, which for entries with {{quote-book}} templates was a
full tablet transliteration (e.g., "o-di-do-si du-ru-to-mo / ..."),
not the headword. This caused 5 entries (i-qo, ma-ra-tu-wo, o-do-ke,
o-di-do-si, wo-ro-ki-jo-ne-jo) to receive IPA from wrong Wiktionary
pages (po-ro, sa-sa-ma, a-re-pa, a-ko-so-ne, e-re-mo).

Fix: prefer unicode_to_translit(title) as primary transliteration,
only fall back to tr= from {{head|gmy|...}} templates. Removes 6
spurious entries created by wrong tr= extraction. Count: 2484 → 2478.

README.md

@@ -150,7 +150,7 @@ The Gothic Bible is the primary source of Gothic text. The paper uses unsegmente
 |---|---|---|
 | `linear_b_signs.tsv` | [Unicode UCD](https://www.unicode.org/Public/UCD/latest/) | 211 signs: 88 syllabograms + 123 ideograms with Bennett numbers and IPA |
 | `sign_to_ipa.json` | Ventris & Chadwick (1973) | 74 syllabogram transliteration → IPA mappings |
-| `linear_b_words.tsv` | Multiple (see below) | 2,484 words with IPA, glosses, and source attribution |
+| `linear_b_words.tsv` | Multiple (see below) | 2,478 words with IPA, glosses, and source attribution |
 
 **Format:** Tab-separated values. The word list contains columns: `Word` (transliteration), `IPA`, `SCA` (sound class), `Source`, `Concept_ID`, `Cognate_Set_ID`, `Gloss`, `Word_Type`, `IPA_Source`. Words come from three CC-BY-SA compatible sources: [jhnwnstd/shannon](https://github.com/jhnwnstd/shannon) Linear B Lexicon (MIT, 2,272 entries), [Wiktionary](https://en.wiktionary.org/wiki/Category:Mycenaean_Greek_lemmas) Mycenaean Greek lemmas (CC-BY-SA, 170 entries with 46 expert IPA), and IE-CoR cognate pairs (42 entries). The sign inventory is from the Unicode Character Database.
 

data/linear_b/README.md

@@ -6,7 +6,7 @@ Structured dataset for Linear B inscriptions and Mycenaean Greek vocabulary, com
 
 - `linear_b_signs.tsv` — Full sign inventory (211 signs: 88 syllabograms + 123 ideograms)
 - `sign_to_ipa.json` — 74 syllabogram transliterations mapped to IPA values
-- `linear_b_words.tsv` — 2,484 words with transliterations, IPA, glosses, and source attribution
+- `linear_b_words.tsv` — 2,478 words with transliterations, IPA, glosses, and source attribution
 - `README.md` — This file
 
 ## Sign Inventory (`linear_b_signs.tsv`)
@@ -38,20 +38,20 @@ Columns: `Word`, `IPA`, `SCA`, `Source`, `Concept_ID`, `Cognate_Set_ID`, `Gloss`
 
 | Metric | Count |
 |--------|-------|
-| Total words | 2,484 |
-| Common nouns | 680 |
+| Total words | 2,478 |
+| Common nouns | 674 |
 | Anthroponyms (personal names) | 1,285 |
 | Toponyms (place names) | 209 |
 | Unknown meaning | ~250 |
 | Theonyms (deity names) | ~27 |
-| Expert IPA (Wiktionary/IE-CoR) | 77 |
-| Transliteration-derived IPA | 2,407 |
+| Expert IPA (Wiktionary/IE-CoR) | 81 |
+| Transliteration-derived IPA | 2,397 |
 
 ### IPA Sources
 
 IPA transcriptions come from three tiers of quality:
-1. **Expert** (77 words): Scholarly reconstructions from Wiktionary `ts=` parameter and IE-CoR cognate pair data. Examples: `a-ke-ro` → `áŋɡelos`, `a-ku-ro` → `árguros`
-2. **Transliteration conversion** (2,407 words): Systematic application of the Ventris grid IPA mapping to each syllable in the transliteration
+1. **Expert** (81 words): Scholarly reconstructions from Wiktionary `ts=` parameter and IE-CoR cognate pair data. Examples: `a-ke-ro` → `áŋɡelos`, `a-ku-ro` → `árguros`
+2. **Transliteration conversion** (2,397 words): Systematic application of the Ventris grid IPA mapping to each syllable in the transliteration
 3. **IE-CoR cognate data** (42 words): 42 Mycenaean Greek words from the IE-CoR (Indo-European Cognate Relations) database with expert IPA reconstructions, linked to Concepticon concept IDs
 
 ### Word Types
@@ -81,7 +81,7 @@ IPA transcriptions come from three tiers of quality:
 
 ## Limitations
 
-1. **IPA quality varies**: Only 77 words have expert IPA reconstructions. The remaining 2,407 use systematic transliteration conversion, which captures the syllabic structure but not phonological details like aspiration, vowel length, or accent.
+1. **IPA quality varies**: Only 81 words have expert IPA reconstructions. The remaining 2,397 use systematic transliteration conversion, which captures the syllabic structure but not phonological details like aspiration, vowel length, or accent.
 2. **Glosses are noisy**: The Shannon lexicon definitions contain scholarly concordance notes, not clean dictionary glosses. Parsing extracts the primary meaning but some entries retain bibliographic noise.
 3. **Name-heavy**: 73% of entries are proper nouns (names, places), reflecting the administrative nature of Linear B tablets (palace inventories, tax records). Only 27% are common vocabulary.
 4. **No tablet corpus**: The dataset includes individual words but not full tablet texts. DAMOS (CC-BY-NC-SA-4.0) and LiBER (non-profit only) provide tablet corpora but are license-incompatible.

data/linear_b/linear_b_words.tsv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d1151d4b9a04b3b3e26b8f5726511b7064b5e495d417c508e20e4abc40c1a3d2
-size 267875
+oid sha256:141571f681c6363c438c2a4a8c8323433f878290f2de176fff7096c462ad3e84
+size 266984

docs/changelog/008_linear_b_dataset.md

@@ -104,15 +104,15 @@ Random sample of 10 entries verified:
 |------|--------:|-----:|
 | `data/linear_b/linear_b_signs.tsv` | 211 | 12,875 bytes |
 | `data/linear_b/sign_to_ipa.json` | 74 | 1,134 bytes |
-| `data/linear_b/linear_b_words.tsv` | 2,484 | 276,697 bytes |
+| `data/linear_b/linear_b_words.tsv` | 2,478 | 266,984 bytes |
 | `data/linear_b/README.md` | — | documentation |
 
 ### Word List Breakdown
 
 | Category | Count | % |
 |----------|------:|--:|
-| Common nouns | 680 | 27.4% |
-| Anthroponyms | 1,285 | 51.7% |
+| Common nouns | 674 | 27.2% |
+| Anthroponyms | 1,285 | 51.9% |
 | Toponyms | 209 | 8.4% |
 | Unknown meaning | ~250 | ~10.1% |
 | Theonyms | ~27 | ~1.1% |
@@ -122,8 +122,8 @@ Random sample of 10 entries verified:
 
 | Source | Count | % |
 |--------|------:|--:|
-| Expert (Wiktionary/IE-CoR) | 77 | 3.1% |
-| Transliteration conversion | 2,407 | 96.9% |
+| Expert (Wiktionary/IE-CoR) | 81 | 3.3% |
+| Transliteration conversion | 2,397 | 96.7% |
 
 ## Limitations
 

docs/changelog/INDEX.md

@@ -6,7 +6,7 @@ All changes to the `Nacryos/ancient-scripts-datasets` HuggingFace dataset are lo
 
 | Date | Entry | Summary |
 |------|-------|---------|
-| 2026-03-19 | [008_linear_b_dataset.md](008_linear_b_dataset.md) | Linear B (Mycenaean Greek) dataset: 211 signs, 74 IPA mappings, 2,484 words from Unicode/Shannon/Wiktionary/IE-CoR |
+| 2026-03-19 | [008_linear_b_dataset.md](008_linear_b_dataset.md) | Linear B (Mycenaean Greek) dataset: 211 signs, 74 IPA mappings, 2,478 words from Unicode/Shannon/Wiktionary/IE-CoR |
 | 2026-03-19 | [007_phono_quality_flagging.md](007_phono_quality_flagging.md) | Added `Phono_Quality` column (strong/moderate/weak/none/unscored), downgraded 27K Robbeets cross-family to "contested", fixed 2.9K Sino-Tibetan doubt markers |
 | 2026-03-19 | [006_tier1_cldf_ingestion.md](006_tier1_cldf_ingestion.md) | +573K expert cognate pairs from IE-CoR, Robbeets, Savelyev — 31 new ancient languages, 4-agent adversarial audit |
 | 2026-03-15 | [005_parquet_conversion.md](005_parquet_conversion.md) | Added Parquet files + YAML dataset card for HF `datasets` library integration |

scripts/build_linear_b_dataset.py

@@ -312,19 +312,29 @@ def parse_wiktionary_lemmas(json_path: Path) -> list[dict]:
             if not tr_check:
                 continue
 
-        # Extract IPA from ts= parameter
+        # Extract IPA from ts= parameter — ONLY from {{head|gmy|...}} template,
+        # NOT from {{quote|gmy|...}} tablet quotation contexts.
+        # The head template has format: {{head|gmy|noun|ts=VALUE}}
+        # Quote templates have format: {{quote|gmy|...|ts=FULL_SENTENCE}}
         ipa = ""
-        ts_match = re.search(r"\|ts=([^|}]+)", wikitext)
-        if ts_match:
-            ipa = ts_match.group(1).strip()
-
-        # Extract transliteration: prefer tr= parameter, fallback to Unicode conversion
-        translit = ""
-        tr_match = re.search(r"\|tr=([^|}]+)", wikitext)
-        if tr_match:
-            translit = tr_match.group(1).strip()
+        head_match = re.search(r"\{\{(?:head|h)\|gmy\|[^}]*\|ts=([^|}]+)", wikitext)
+        if head_match:
+            ipa = head_match.group(1).strip()
+            # Sanity check: headword IPA should be a single word, not a sentence
+            # If it contains spaces or <br>, it's a tablet quotation that leaked in
+            if " " in ipa or "<br>" in ipa:
+                ipa = ""
+
+        # Extract transliteration: prefer Unicode title conversion, fallback to tr= from head template.
+        # IMPORTANT: Do NOT use a global tr= search — {{quote-book}} templates contain
+        # tr= with full tablet transliterations (e.g., "o-di-do-si du-ru-to-mo / ..."),
+        # which are NOT the headword transliteration. Only use tr= from {{head|gmy|...}}.
+        translit = title_translit  # Primary: Unicode character names → transliteration
         if not translit:
-            translit = title_translit
+            # Fallback: tr= from head template only
+            head_tr_match = re.search(r"\{\{(?:head|h)\|gmy\|[^}]*\|tr=([^|}]+)", wikitext)
+            if head_tr_match:
+                translit = head_tr_match.group(1).strip()
 
         # Clean transliteration: remove tablet context, bold markers, etc.
         # Wiktionary titles sometimes embed context like "'''di-wo''' u-ta-jo-jo"