Release 7.51.1a1 - #778
Open
github-actions[bot] wants to merge 199 commits into
Open
Conversation
* Update kab.json Add post-sonorant occlusivization blocking rules for /k/ and /ɡ/ * kab: rework k/ɡ spirantization, fix Bedar et al. citation kab: rework k/ɡ spirantization, fix Bedar et al. citation - Remove KAB_SPIRANT_K_ALT and the [ʃ] allophone of /k/: the k→ʃ value is Riffian-only in Kossmann & Stroomer (1997) p.469 and never occurs in Kabyle (native-speaker verified). - Replace the 7 sonorant-specific blocking rules (K/G_AFTER_R/L/N/M) with generalized KAB_BLOCK_SPIRANT_K_AFTER_CONSONANT and KAB_BLOCK_SPIRANT_G_AFTER_CONSONANT: /k/→[ç] and /ɡ/→[ʝ] only post-vocalically and word-initially; stops after any consonant. Exception kept: /ɡ/ stays [ʝ] after /m/ (/mʝ/→[mʝ], Bedar, Quellec & Tifrit 2022). Contexts cover all singleton + geminate consonants. * kab: drop t-before-i block, restore anta example, add 6 sources kab: drop t-before-i block, restore anta example, add 6 sources - Remove KAB_BLOCK_SPIRANT_T_BEFORE_I: speaker re-verification (2026-07-17) confirms word-initial /t/ is always [θ] before /i/ (tifinaɣ, tilufa, tilelli, tikwal) — the 2026-07-16 table stands. 24 → 23 allophone rules. - Restore anta as the T_AFTER_N example with corrected gloss: native Kabyle 'which one (f.)' (feminine of anwa), not the Arabic pronoun — only the old gloss was wrong. - Sources 6 → 12: add Chaker 1996 (INALCO notation), Mammeri 1976, Souag 2019, Wang 2020 (McGill Working Papers — corrects a miscitation), the 2026-07 native-speaker verification record, and the 2026-07-17 corpus audit. Notes now list the pending speaker checks (intervocalic -ti-; askar/sk schwa question). Engine-verified on orthography2ipa 7.30.0a3 (fork): all outputs match predictions. * kab: realize /ɣ/ as uvular [ʁ] by default kab: realize /ɣ/ as uvular [ʁ] by default Source check (Kabyle consonant chart: ɣ in the uvular column as [ʁ]; IPA [ʁ] article's word-final Kabyle example bbeɣ → [bːəʁ]; no word-final devoicing documented for Kabyle): add KAB_GH_UVULAR_REALIZATION and its geminate counterpart so /ɣ/ surfaces as [ʁ] in all positions. Underlying phoneme and rule contexts unchanged; broad [ɣ] retained as documented variant. 23 → 25 rules. * Improving kabyle * Improving kabyle specs * Improved kabyle specs * Improved kabyle specs - json format
…greement boards (#636) The scoreboard now carries 553 rows across every gold set the engine ships, up from 467: the languages added since the last regeneration are scored, cmudict and the Arabic diacritized set included. The espeak board gains en-US and splits the single pt row into the five Portuguese lects.
…d a stale regression baseline (#637) * fix(kab): validate against the schema, and re-pin the gold to the sourced rules The Kabyle spec carried a 'type' key on seven sources[] entries. The schema's SourceModel is strict and has no such field, so every spec validation failed; the provenance claim now opens each source's notes, which the model does define. The spec's rules are better sourced than the expectations they broke: /ʝ/ occlusivizes after r, l, n but not after m (bedar_quellec_tifrit_2022 on Chemini Kabyle), so argaz keeps its stop; the /k/ -> [ʃ] rule was Kossmann & Stroomer's RIFFIAN illustration and its removal makes amek agree with its own cited note ([ç]); and /ɣ/ is the uvular [ʁ] in every position. The tests now assert the sourced behaviour on both sides of the blocking context, the twenty gold rows are re-pinned, and the notes that still described the old realizations are corrected. The velar_fricative tag accepts either realization of the one phoneme, and kab-004 drops a spirantization tag its transcription no longer earns (fourteen other rows carry the feature). * fix(ci): install the Portuguese syllabifier in the test environment The gold sets and benchmark baselines pin engine output produced with the silabificador plugin loaded, but the plugin lived only in the 'portuguese' extra while build-tests, coverage and the benchmark gate all install '.[test]'. Without it the syllabifier falls back and stress lands elsewhere, so twenty Portuguese gold rows across fourteen lects failed in CI while passing locally for anyone who had the plugin installed. * fix(ci): refresh the benchmark regression baseline The sample baseline had not been refreshed since #361, so it covered 206 of the 551 rows the engine now scores and measured dev's whole accumulated drift rather than a PR's own effect. Every PR since — #614 among them — met a red gate and merged through it, which is what a baseline this old makes inevitable. Rescored at the same uniform limit: 56 rows improve beyond epsilon (ms/ipadict 0.5509 -> 0.0751, kab/vox_communis 0.3711 -> 0.2242, eu/wikipron 0.0640 -> 0.0227) and 8 worsen. The 8 are listed in the PR so they stay visible rather than folding silently into a new baseline; ca-x-balear/4catac and pt-PT-x-porto/ep_dialects are the two worth a look, both expert-human golds moved by the Iberian vocalism work.
Expose the prosodic-clitic test as module-level orthography2ipa.stress functions so the engine and any downstream assembler place word stress from one shared definition of stress.cliticless_words. G2P._is_cliticless delegates to them; behaviour is unchanged. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
A bare region tag whose specs are all sub-regional had no exact match, so nearest-language fallback picked whichever private-use sibling sorted first: ar-SA resolved to ar-SA-x-dawasir. A dialect named in a private-use subtag was invisible to tag distance, so ar-x-najdi fell through to ar. Add a curated region-default table (ar-SA -> ar-SA-x-najd) alongside the bare-primary defaults, and private-use dialect aliases (ar-x-najdi -> ar-SA-x-najd, ar-x-hejazi -> ar-SA-x-hejaz). The Saudi default is an explicit editorial convention — Najdi is the capital region's variety and the most widely spoken (Ingham 1994) — not a claim in the spec data that Saudi Arabic is Najdi. Unmatched Arabic tags still fall back to the ar (MSA) leaf. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The grapheme→IPA front-end runs once per utterance before synthesis, so per-sentence latency is what matters. Profiling the transcribe path showed two pure redundancies over immutable spec data: - transcribe_word ran the pronunciation beam, then rebuilt a nearly identical rules-only lattice via ipa_lattice purely to read the confidence score. - syllabification (the silabificador plugin) was recomputed several times per word — stress detection and the lattice each asked for it — and again for every repeat of a function word. Memoize both per engine, keyed only on the word and this engine's fixed stress spec. Neither consults the runtime-mutable lexicon, so a fresh engine gets fresh caches and results are byte-identical. Real speech repeats function words heavily, so the hit rate is high. Steady-state per-utterance latency drops ~2-2.6x across en-GB, pt-PT, es-ES and ar (pt-PT p99 7.2ms -> 3.4ms). Output — IPA and confidence, greedy and beam — is byte-identical across an 8-language differential, and the full suite passes unchanged. benchmarks/latency.py is the repro. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…#654) * feat: add gaa, ada, gen, aja language specs Ga, Dangme/Adangme, Gen (Mina) and Aja — Ghana/Togo Kwa and Gbe languages. Research-tier grapheme/allophone inventories cited to Wikipedia and published phonological descriptions (Kropp Dakubu 1999, Apronti 1967); gen/aja pull their allophone stratum from the closely related Ewe (`ee`) spec via allophones_base. * feat: add tw, fat Akan-variety language specs Twi and Fante as individual-language codes inheriting the shared Akan (`ak`) grapheme table via graphemes_base/allophones_base; fat overrides Fante's distinguishing affricate contrast (ts/dz/z) and labial-palatalised digraphs (tw/dw/hw). Pins the two individual codes in registry.py's alias table (same fix already applied to bxr/diq) since langcodes' macro=True standardisation otherwise collapses them into the ak macrolanguage. Refreshes the Glottolog level snapshot for the five new glottolog_code values and adds cited family-level test cases for all six new specs.
…ulde specs (#655) * feat: add Serer (srr) language spec Northern Atlantic (Senegambian) branch of Atlantic-Congo. Cited to Wikipedia and McLaughlin (2000); quality: research. * feat: add Soninke (snk) language spec Soninke-Bozo branch of Mande. Cited to Wikipedia and Creissels (2016); quality: research. * feat: add Susu (sus) language spec Susu-Yalunka branch of Central Mande. Cited to Wikipedia and Houis (1963); quality: research. * feat: add Mende (men) language spec Southwestern Mande. Cited to Wikipedia and Innes (1969); quality: research. * feat: add Temne (tem) language spec Mel branch of Atlantic-Congo. Cited to Wikipedia and Wilson (1961); quality: research. * feat: add Pular (fuf) and Nigerian Fulfulde (fuv) language specs Both wired as parent_dialect children of ff via graphemes_base/parent, since neither source consulted documents a Fula-variety-specific grapheme or allophone divergence. Cited to Wikipedia and Arnott (1970); quality: research. * test: add cited-rule tests for the new Atlantic and Mande specs One test per key cited claim (srr vowel length + implosives, snk vowel length + affricate, sus/men labiovelars, men affricate vs. sus glide, tem dental/alveolar contrast, fuf/fuv ff inheritance) on real words. * fix(data): correct citation-validation findings for sus and tem specs - sus.json: Wikipedia's Susu phoneme table (fetched) explicitly states nasal vowels are spelled as vowel+n digraphs (in/un/en/on/ɛn/ɔn/an); the previous notes claimed this was unconfirmed and omitted the graphemes. Add the seven nasal-vowel graphemes/allophones and fix the notes to cite the confirmed spelling convention. - tem.json: add Kanu & Tucker (2010) JIPA "Illustrations of the IPA" as the primary fetched source, confirming the 9-vowel inventory and the dental/alveolar ⟨th⟩=/t̪/ vs ⟨t⟩=/t/ contrast, and clarifying that /tʃ/, /ʃ/, /w/ are dialectal variants rather than independently spelled phonemes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The README and several docs quoted a fixed 676 languages / 749 spec files figure. Running available_codes() against the current data now gives 820 languages and 909 entries including classification-only clade nodes, and that count changes as specs are added. Rather than hardcode a new snapshot that will drift the same way, each mention now points at the one-line command a reader can run to get the live count: available_codes() and available_codes(include_clades=True). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: add Sasak, Rejang, Pijin language specs
Research-grounded Latin-orthography specs for Sasak (sas, Lombok),
Rejang (rej, Sumatra) and Pijin (pis, Solomon Islands), each cited
against Wikipedia and published grammars/phonology sources.
* feat: add Avestan spec and Sogdian/Khotanese/Parthian stubs
Avestan (ae) is scored in the standard Hoffmann scholarly
transliteration of the native Avestan script, following the akk.json/
pal.json precedent for ancient languages with non-Latin scripts.
Sogdian (sog), Khotanese (kho) and Parthian (xpr) are added as honest
stubs with full metadata and empty graphemes: each survives in
multiple divergent script/transliteration traditions with no single
citable letter-to-IPA table, so no inventory is asserted.
* fix(data): validate sas/rej/pis/ae specs against fetched primary sources
- sas.json: add Archangeli/Tanashur/Yip (2020) JIPA Illustration of the
Meno-Mene dialect as primary source; model word-final <q> = /ʔ/ as a
positional_graphemes rule (directly attested in the JIPA word list);
clarify the 6-vowel (JIPA fieldwork) vs 8-vowel (Wikipedia, other
dialects) vowel-count discrepancy in notes.
- rej.json: add fetched Omniglot citation; model word-final <k> = [ʔ]
as a positional_graphemes rule ("K is pronounced [ʔ] at the end of
words" - Omniglot); leave <q> unmodelled since no IPA value is
confidently sourced for it.
- pis.json: remove unsupported <y> = /j/ grapheme (Wikipedia's Pijin
alphabet excludes y; search evidence points to <i> being used
instead, e.g. "iu" for "you"); fix notes claim that both voiceless
and voiced English th merge with /t/ - fetched Wikipedia source shows
voiceless th merges with /s/ (maos "mouth"), only voiced th merges
with /t/ (brata "brother"); add sh/z merger details.
- ae.json: add Skjaervo (2007) "Old Iranian" chapter (fetched directly,
PDF) as the actually-verified primary source after the Encyclopaedia
Iranica articles by the same author returned HTTP 403 on every fetch
attempt; cross-checked the full Hoffmann transliteration letter set
against Skjaervo's own alphabet/vowel/consonant tables - no mapping
corrections were needed, all graphemes confirmed.
All four specs pass the full test suite (1208 passed, 2 pre-existing
xfails unrelated to these files) and the derivable-key/positional-key
audits from SKILLS.md.
…drift (#670) The CUNY-CL/wikipron Welsh scrape changed upstream (897 -> 894 entries), shifting the sampled gold set and its PER (0.2175 -> 0.2297) with no code or spec change in this repo — verified by bisecting dev back past all of today's data merges and the #645 caching merge, which all score identically. Rebuild the single (cy, wikipron) row with the same build_scoreboard code path (bootstrap CI included), following the stale baseline precedent of #637. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Citation-verifies 6 more skeleton-tier specs from the fresh 37-candidate pool (skeleton + non-empty graphemes + registered gold, per docs/quality_tiers.md's research-tier criteria): kru, bbl, kwk, srs, kld, bdq. Each already had a real Wikipedia orthography source; the gap was a primary reference grammar/dictionary and a cited stress block/exemption. - kru (Kurukh): added Winfield 1928 (A Grammar of the Kurux Language). Grapheme table (Devanagari retroflex/aspirate/nasal series, nuqta letters) checked: no discrepancy found. Descriptive phonological work reports Kurukh word accent as weak and variable across syllables rather than a fixed positional rule, so no `stress` block is added (documented exemption) rather than encoding a guess. - bbl (Bats/Tsova-Tush): added Holisky & Gagua 1994 (Tsova-Tush (Batsbi), in Smeets ed., Indigenous Languages of the Caucasus vol. 4). Grapheme table (Georgian ejective series, added letters) checked: no discrepancy found. Added a cited stress rule (default_position=1) approximating the documented initial-syllable tendency; morphologically-conditioned shifts are a documented gap. - kwk (Kwak'wala): added Boas 1947 (Kwakiutl Grammar, Transactions of the American Philosophical Society 37(3)). Grapheme table (labialized/lateral/affricate series) checked: no discrepancy found. Stress is quantity-sensitive and, in current analyses, phrasal rather than a fixed word position (Boas 1947; Elfner 2024), which this project's positional `stress` schema cannot express without a syllable-weight primitive, so no `stress` block is added (documented exemption). - srs (Sarsi/Tsuut'ina): added Cook 1984 (A Sarcee Grammar). Grapheme table (lateral/ejective/affricate series) checked: no discrepancy found. Tsuut'ina is a tone language (Cook 1984; reported as a three-way H/M/L system in later work), not a stress language, and tone is not consistently marked in the mapped orthography, so no `stress` block is added (documented exemption). - kld (Gamilaraay): added Austin 1993 (A Reference Grammar of Gamilaraay, Northern New South Wales). Grapheme table (stop series, retroflex r, digraph values) checked: no discrepancy found. Added a cited stress rule (default_position=1): word-initial primary stress, the typologically regular Pama-Nyungan pattern. - bdq (Bahnar): added Banker & Banker 1979 (Bahnar Dictionary, Bahnar-Vietnamese-English, SIL). Grapheme table (Vietnamese-style digraphs, implosives, vowel letters) checked: no discrepancy found. Added a cited stress rule (default_position=-1): Bahnaric sesquisyllabic words carry stress on the word-final main syllable; non-native disyllabic loans are a documented gap. None of the six required a grapheme-table fix, so no PER numbers changed; only each spec's `quality_tier` was hand-updated in benchmarks/results.json and docs/scoreboard.md (not regenerated via --scoreboard/--ci-sample), per wave-2 precedent. No CI-sample epsilon concerns are expected since no PER value moved. 31 candidates remain in the fresh audit pool (skeleton + non-empty graphemes + registered gold, missing a real source or stress block/exemption) after this wave's 6 promotions, left for a future wave. `tn` and `ale` were left untouched: `tn`'s only registered gold row is a 0-entry stub-tier placeholder (no usable PER to certify a promotion against), and `ale` carries a prior deliberate skeleton decision (documented gold mismatch) that this wave respects rather than re-litigating. Authorship: Claude Sonnet, orchestrated by Fable, per repo AGENTS.md. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Gold-hunting wave 1: audited all 41 WOLD (World Loanword Database) languages against the o2i spec registry. Gawwada, Iraqw, Seychelles Creole and Tarifiyt Berber each have a non-empty grapheme table and no existing gold anywhere in the registry; all four smoke-checked at ~100% non-empty engine coverage on a 150-row sample. Rows recorded in benchmarks/results.json and docs/scoreboard.md, with the full audit (what was already covered, what has no spec, what is stub-empty) documented in docs/benchmarks.md. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* data: orthography wave 6 -- 10 registry stubs upgraded with cited orthographies Cantonese (yue, Jyutping), Saraiki (skr, Shahmukhi) promoted to research tier and wired to WikiPron gold (yue_latn_standard_broad, skr_arab_broad). Southern Min (nan, POJ), Hakka (hak, Phåk-fa-sṳ), Wu (wuu, Wugniu), Eastern Min (cdo, Bàng-uâ-cê), Gan (gan, Gan Wikipedia romanisation), Northern Hindko (hno, Shahmukhi), Rohingya (rhg, RLC Latin alphabet) and Chittagonian (ctg, Bengali script with cited divergent values) promoted to skeleton tier with cited grapheme maps and documented tone/stress exemptions. Co-Authored-By: Claude Sonnet 5 (Fable orchestration) <noreply@anthropic.com> * fix: drop Sinitic base-spec romanization graphemes and the loanword yue gold The five Han-script base specs (yue/nan/hak/wuu/cdo) must stay grapheme- empty per the zh-Hani principle: a Han character encodes no sound, and their romanizations already live in the dedicated -Latn-* script-variant specs (which these divergent duplicate tables contradicted). gan's 'Gan Wikipedia romanisation' is a wiki convention, not a citable standard. The yue wikipron file was previously audited and rejected as English loanword entries (PER 0.71 confirms); registration and rows removed. Wave 6 keeps its four legitimate native-script languages: skr Saraiki (Shahmukhi, wikipron research), hno Hindko, rhg Rohingya, ctg Chittagonian. --------- Co-authored-by: Claude Sonnet 5 (Fable orchestration) <noreply@anthropic.com>
Wires udm, ady, av, lbe, dar, lez, lv, smn, vep into _NORTHEURALEX_LANGS: each had zero gold anywhere and a non-empty grapheme table. Language-ID mappings were verified against NEL's own cldf/languages.csv (not a naive ISO 639-3 lookup) -- av -> ava and lv -> lav needed an explicit code translation, and dar was checked against its glottocode (darg1241, "North-Central Dargwa") to confirm it is the same Akusha-based literary standard the o2i dar spec targets. All 9 passed a 100-sample smoke gate (100% non-empty output, real per-language orthography in Value) before full-dataset scoring. dar is promoted skeleton -> research: it already had sources and a documented stress exemption, and now has a registered gold row (1056 entries, PER 0.0833) to complete the tier's criteria. The other 8 languages were already labeled research pending exactly this gold registration. benchmarks/results.json and docs/scoreboard.md are hand-updated with the new rows (full-dataset PER, no --scoreboard/--ci-sample rerun). Authorship: Claude Sonnet, orchestrated by Fable (Anthropic). Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…ve-speaker spot-checked (#767) Registers Salesteq/arabic-dialects-gold20 (public HF dataset) at the llm-generated provenance tier: semi-synthetic, authored by the same Claude lineage that wrote o2i's Arabic dialect specs (near-circular — the dataset carries an ipa_o2i draft column), spot-checked by a native Arabic speaker, registered on explicit maintainer instruction because for most of these 33 dialects no other gold exists. PERs 0.002-0.05 = self-agreement, documented as such; no quality-tier promotion may rest on these rows. Runtime fetch only, nothing vendored.
* feat: gold-hunting wave 2 -- register kaikki.org Wiktextract gold for 4 zero-gold languages Wires a new generic kaikki.org (Wiktextract) loader and registers gold for jv (Javanese), su (Sundanese), lo (Lao) and xh (Xhosa) -- all zero-gold skeleton specs with a non-empty grapheme table, cross-referenced against the full o2i registry before selection. Each language was downloaded, filtered to sounds[].ipa entries, hand-sampled, and smoke-checked for >=70% non-empty engine coverage; jv needed a Latin-script filter since kaikki's Javanese dump is majority Aksara Jawa. Tigrinya was rejected (only 28/933 entries carry ipa). Findings on lo/xh spec gaps recorded in docs/benchmarks.md for a future (non-gold) wave. No promotions: all four lack a stress block/exemption required for the research tier. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * chore: record the four kaikki rows in results.json + scoreboard --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
…771) Promote 6 empty REGISTRY STUB specs (ckt Chukchi, itl Itelmen, ket Ket, gld Nanai, ddo Tsez, ess Central Siberian Yupik) from stub to research tier with cited Cyrillic grapheme tables, NorthEuraLex gold registered and PER recorded after one error-analysis iteration each. mnc (Manchu) is promoted using the Moellendorff Latin romanization, verified as the actual orthographic convention NorthEuraLex's Value column uses (not the vertical Manchu script), cited to Moellendorff 1892 / Gorelova 2002. bsk (Burushaski) is evaluated and deliberately NOT promoted: its NEL gold is Berger (1998) scholarly transcription, not a community orthography, so no graphemes or gold registration are added -- the verdict is documented in the spec's notes. Authorship note: drafted by Claude (Sonnet) under Fable orchestration.
* data: NEL convention fixes — fold ASCII/IPA g confusable in the harness The 9 newly-measured NorthEuraLex languages from #765 (udm, ady, av, lbe, dar, lez, lv, smn, vep) all showed a `g -> ɡ` confusion in scripts/error_analysis.py: NEL's CLDF Segments column keys the voiced velar plosive with ASCII "g" (U+0067) while every o2i spec correctly emits the IPA symbol ɡ (U+0261, LATIN SMALL LETTER SCRIPT G). No registered spec's own phonemes list contrasts the two (checked against every data/*.json), so this is a Unicode-confusable typographic accident in the gold data, not a phonemic distinction — folded unconditionally in scripts/benchmark.py's normalize(). PER before -> after (northeuralex, full dataset): udm 0.2287 -> 0.2110 ady 0.1745 -> 0.1684 av 0.1115 -> 0.0990 lbe 0.0443 -> 0.0387 dar 0.0833 -> 0.0691 lez 0.1513 -> 0.1329 lv 0.2046 -> 0.1901 smn 0.1755 -> 0.1600 vep 0.2276 -> 0.2065 Other seeded udm confusions (gold ɑ vs spec a, gold l vs spec ɫ, gold sʲ/zʲ/lʲ vs spec ɕ/ʑ/ʎ palatalization notation) were investigated and NOT changed: several other registered specs contrast both members of each pair in their own phoneme inventories (e.g. tyv/mzn contrast ɑ/a, sjt/enf contrast sʲ/ɕ, diq/kdr contrast ɫ/l), so a normalizer fold is not linguistically defensible for ALL languages — these are legitimate per-language transcription-convention differences, not spec bugs, and are left alone per the "don't chase" guidance. Because the fold lives in the shared harness (normalize()), it also shifts PER for the co-registered NEL Siberian-batch codes (liv/sms/sjd/ yrk/bua/evn/niv/ale/ain — verified liv 0.1019 -> 0.0748) and possibly other g-bearing rows outside northeuralex; those rows are intentionally left untouched here (out of scope, owned by other in-flight work) for the next full scoreboard regen to pick up. Co-authored-by: Claude Sonnet under Fable orchestration * chore: refresh the 9 co-registered NEL rows for the g/ɡ fold * fix: fully recompute the 8 co-affected NEL rows (per+CI+exact_match) — adversarial review finding
* data: Siberian double-win batch A (Uralic+Yukaghir) Cited Cyrillic grapheme tables for 8 REGISTRY STUBs whose NorthEuraLex gold was already waiting, promoting each to research tier: Meadow Mari (mhr, Alhoniemi 1985), Hill Mari (mrj, Kangasmaa-Minn 1998), Northern Khanty (kca, Nikolaeva 1999), Northern Mansi (mns, Riese 2001), Northern Selkup (sel, Helimski 1998 in Abondolo), Nganasan (nio, Wagner-Nagy 2018), Forest Enets (enf, Siegl 2013) and Northern Yukaghir (ykg, Maslova 2003). Each spec reconciles its grapheme table against the pre-existing PHOIBLE phoneme inventory, documents a cited stress exemption, registers the NorthEuraLex gold row (Language_ID verified against cldf/languages.csv; all eight equal the o2i code, no av->ava-style translation needed), and went through one error-analysis iteration against the gold before the final PER was recorded: - mhr: 0.3037 -> 0.1869 (a as /ɑ/, л palatalises to fused /ʎ/) - mrj: 0.0629 -> 0.0140 (д as lenited /ð/, not Meadow Mari's /d/) - kca: 0.3236 -> 0.0232 (ӆ as /ɬ/, х as /χ/, ы as /ɪ/, э as /ə/, т unaspirated) - mns: 0.3102 -> 0.1024 (vowel-quality fixes, labialised кв/хв clusters, plain ɲ/ʎ palatals, sʲ series added) - sel: 0.1852 -> 0.0353 (a not ä, fused ɲ/ʎ palatals) - nio: 0.2283 -> 0.1075 (a as /ɐ/, complete н/д palatalisation series) - enf: 0.3161 -> 0.1073 (a not ɑ, е/э as /ɛ/, х as /h/, ɲ/ʎ palatals, ⟨ҫ⟩ added) - ykg: 0.5090 -> 0.3940 (ч as /tʃ/, нь/дь as /nʲ/dʲ/ per Maslova's own transcription convention, preferred over inherited PHOIBLE symbols) Authorship: Claude Sonnet under Fable orchestration. * fix: kca same-corpus-tuning caveat + mns fixed-initial stress block — adversarial review findings
* fix: n=0 gold vox_communis rows (tn, ug fake-stub; yue dead pairing)
benchmarks/results.json recorded tn/ug/yue vox_communis rows at a fake
per=1.0/n=0 ("stub" tier snapshot). n=0 in build_scoreboard() means
`covered` (G2P-transcribable words) was zero, not that the loader
returned nothing -- the upstream tn.tsv/ug.tsv/yue.tsv files all fetch
fine (live-checked) and load_vox_communis() yields thousands of pairs
for each.
Root cause differs per language:
- tn, ug: the recorded n=0 was STALE. Both specs now transcribe their
respective gold real words fine (tn: 2271/2272 covered, PER 0.40;
ug: 40716/40721 covered, PER 0.09) -- re-running the same scoring
code the committed row used shows they never should have been 0.
Refreshed both rows with current numbers.
- yue: genuinely UNTRANSCRIBABLE, not stale. The `yue` (Cantonese)
spec is a documented grapheme-inventory STUB: Cantonese is
logographic and has no letter-to-sound mapping without a
Jyutping/Yale romanisation step this library doesn't have (same
class of gap as `zh` needing pinyin). vox-communis' yue.tsv gives
raw Han characters, so G2P('yue') emits "" for every row -- the
n=0/per=1.0 the harness recorded was real, but reporting it as a
benchmark result is dishonest (reads as "loader broken", not "this
pairing can never score"). Deregistered `yue` from
`_VOX_COMMUNIS_FILES` with a comment explaining why, matching the
existing `_IPADICT_UNWIRED`-style honesty precedent for the same
Han-script problem, and dropped its row from results.json/scoreboard.md.
Full test suite green (196593 passed, 0 failed).
Co-Authored-By: Claude Sonnet 5 (Fable) <noreply@anthropic.com>
* data: sources wave 4 -- 2 skeletons promoted to research (tn, chb)
Fresh audit of the skeleton + non-empty-graphemes + registered-gold
pool (docs/quality_tiers.md research-tier criteria) after wave 3.
Citation-verifies 2 skeletons; the rest of the pool (33 candidates:
ale, aot, apw, bjb, cnk, gwc, gwd, hts, irk, kix, lou, lsi, mtq, ngh,
nhg, nhx, nmy, pbv, pcc, pt-CV, pt-MO, rgn, sce, sdc, sia, tew, twf,
ulw, wbk, xsl, ycl, yux, zom) is skipped honestly this wave: most only
carry Glottolog/Wikipedia/Omniglot citations with no independently
verifiable primary reference behind the grapheme table, and two
(pt-CV, pt-MO, ale) already carry an explicit "kept deliberately at
skeleton, do not promote" note from a prior pass -- respected here,
not re-litigated.
- tn (Tswana): already cited Cole 1955 ("An Introduction to Tswana
Grammar") as a read primary alongside Wikipedia. Grapheme table
(ejective k/p/t/ts/tl/pj vs aspirate kh/ph/th/tsh/tlh series,
implosive b, kg=/x/, g=/ɣ/, click q/qh restricted to
ideophones/loans) checked against it and against the cited
Wikipedia 'Tswana phonology' article: no discrepancy found. Stress
is non-contrastive/predictable in Tswana (tone carries the
distinctive prosodic load instead), documented as an exemption
matching the sibling `st`/Southern Sotho precedent, which was
promoted without a stress block on the same grounds. Gold:
vox_communis, PER 0.40, n=2271 (also fixed in the prior commit).
- chb (Chibcha/Muisca, extinct): the spec's own notes already named
Constenla (1984) and Gonzalez de Perez (2006) in prose as the basis
for the reconstructed phonemic values, but neither was a formal
`sources` entry -- only Glottolog+Wikipedia were. Verified both are
real, independently-published works (Constenla Umaña, "Los fonemas
del muisca", Estudios de Lingüística Chibcha 2:65-111, 1984;
González de Pérez, "Aproximación al sistema fonético-fonológico de
la lengua muisca", Instituto Caro y Cuervo, 2006) and added them as
proper cited sources. Stress is exempt: the language is extinct with
no living pronunciation and neither reconstruction recovers word
stress from the colonial spelling. Gold: wikipron, PER 0.13, n=99.
Full test suite green (196593 passed, 0 failed).
Co-Authored-By: Claude Sonnet 5 (Fable) <noreply@anthropic.com>
* fix: refuse to record zero-coverage scoreboard rows — adversarial review finding
evaluate_words defaults per/wer to 1.0 on covered==0 and build_scoreboard
appended the row unconditionally, so a transient loader failure fabricated
a plausible-looking per=1.0, n=0 row (the exact origin of the stale tn/ug/
yue rows this PR cleans up). build_scoreboard now loudly refuses such rows;
regression tests pin both the refusal and the normal path.
---------
Co-authored-by: Claude Sonnet 5 (Fable) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Human review requested!