lexeme-alignments — surface → original-language lexeme (Strong's-bridged)

For each language, the attested mapping from target surface word-forms → the original-language lexeme they render, mined by the aligner. Lexeme-anchored, provenance-honest, additive — the design principles are in docs/publishing-principles.md. One language per partition, for consumption by bcv-commons and downstream tools.

The language: list above tracks the published partitions; the authoritative list is always manifest.json.

The anchor: lexeme, not Strong's

The anchor of record is the MACULA lexeme (hbo:0430, grc:2316) — the precise dictionary unit. The bare Strong's number is coarser (it conflates homonyms and sense-splits — one Strong's rolls up several lexemes) and is a pure, lossless function of lexeme — so it is not stored (dropped 2026-07: ~32% smaller Parquet, zero information lost). Derive it yourself (or use scripts/strongs_view.py, below) — one exception below (hebrew_lexeme_strong.json) overrides the mechanical derivation for a small, verified set of lexemes:

def strong_of(lexeme: str) -> str:
    lang, num = lexeme.split(":", 1)
    digits = "".join(c for c in num if c.isdigit())          # strip any trailing augment letter
    return ("H" if lang == "hbo" else "G") + digits.zfill(4)  # e.g. "hbo:6498a" -> "H6498"

Schema (per row)

column	type	meaning
surface	string	target rendering, lowercased (content tokens; may be multi-word)
lexeme	string	the anchor — MACULA lexical id (lang:augmented-strong)
method	string	which method attested this pair — eflomal / gloss / gapfill
base_text	string	which edition the surface is from (e.g. BSB, eng_ylt)
count	int32	times this (surface → lexeme) was aligned in that method + edition
hi_conf	float32	fraction of this pair's alignments that were intersection-backed (score ≥ 0.9)

iso is recovered from the Hive partition path (iso=<iso>/). Two honest provenance axes: method (how aligned) and base_text (which edition).

Not stored, both exact + lossless from the columns above (measured: dropping them shrinks the Parquet ~32% with zero information loss — the two derivations are independent, so drop either or both):

• strong — see above.
• share (P(lexeme|surface) within a (surface, method, base_text) group) — group rows by (surface, method, base_text), sum their count, then share = count / that sum.

It's an additive union — nothing is merged away

Rows are the union of the methods, each tagged with its method. A surface→lexeme attested by both eflomal and gloss is two rows (eflomal ×N, gloss ×M) — full provenance, no winner-take-all merge. This means:

• a gapfill-only fact says method=gapfill — it can never masquerade as eflomal/gloss-attested (gapfill is the lower-confidence coverage layer — model-free priors filling positions eflomal+gloss left uncovered; see docs/publishing-principles.md §3 for why this provenance is never hidden);
• an enhanced translation that renders one lexeme with many words keeps all of them — we never force a lexeme to a single "canonical" surface;
• counts are per-method, so do not sum across methods to get an occurrence total (the same verse is often aligned by more than one method — that would double-count; see the on-ramp script).

Cross-method agreement (a real confidence signal, same shape as cross-edition agreement below) — group rows by (surface, lexeme, base_text) and count the distinct method values present. A fact independently found by both eflomal and gloss (two structurally different methods — one statistical, one dictionary-based) is stronger evidence than either alone. Concrete example from the published fra partition: (surface="a", lexeme="grc:2192", base_text="fraLSG") has both a method=gloss row (count=125) and a method=eflomal row (count=102) — two independent methods, same conclusion.

Multiple editions of one language (pooling)

Some languages ship several editions pooled into one partition, each row tagged by base_text (e.g. eng = BSB + YLT; arb = Van Dyck + New Arabic Version; swe = Folkbibeln + Kärnbibeln). This is additive — every edition's renderings are kept, and:

• single edition → filter base_text = '<edition>';
• cross-edition agreement (a strong confidence signal) → a surface→lexeme attested by more than one base_text is corroborated across independent translations; derive it by counting distinct base_text per (surface, lexeme). An enhanced/literal edition (e.g. YLT's begat/begotten) contributes its own renderings without overwriting the others.
• takedown → if a rights-holder objects, drop that base_text's rows and republish (content-addressed); never re-emit provenance-stripped.

The manifest.json entry lists the pooled base_texts, per-edition row counts (by_base_text), and a sources pointer per edition.

Using the data — pick your operating point

Three independent signals; combine them. The dataset ships the full distribution rather than pre-filtering, so precision / coverage / provenance are sliders you control:

goal	filter
everything / max recall	all rows
exclude the gapfill coverage layer	method != 'gapfill'
one edition only	base_text == '<edition>'
cross-edition-corroborated	keep (surface, lexeme) with ≥2 distinct base_text
cross-method-corroborated	keep (surface, lexeme, base_text) with ≥2 distinct method
balanced (recommended default)	argmax-derived-share per (surface, method), count ≥ 2
high precision	hi_conf ≥ 0.5, count ≥ 2
one method only	method == 'eflomal' (or gloss)
treat with extra caution	lexeme is a key in light_lexemes.json (below) — see that section

• share (derived, see above) → which lexeme (P(lexeme|surface) within a method).
• hi_conf → how reliable the placement (intersection-backed share).
• count → how much evidence (count: 1 rests on a single occurrence).
• no recommended universal minimum — count: 1 rows are real (not noise-filtered away), just weaker evidence; if you need a floor, count ≥ 2 is the ablation-tested "balanced" default above.

Derived views (example scripts, never a second source of truth)

1. Strong's on-ramp — roll lexeme→strong from a single base method into a clean Strong's-keyed table (surface + frequency), for ecosystem tools:

   python3 scripts/strongs_view.py --iso swe                    # → out/strongs_view_swe.tsv
   python3 scripts/strongs_view.py --iso swe --hi-conf 0.5 --min-share 0.02
   

   It picks one base method (default eflomal) so per-method counts don't double-count, derives strong from lexeme (checking hebrew_lexeme_strong.json first — see below), then aggregates per (strong, surface) with share = P(surface | strong).
2. Merged best-pick (optional, lossy, NOT reproducible from this dataset alone) — a single-answer- per-token convenience our own pipeline builds from the per-occurrence jsonl (not published — only the aggregated rows here are), using contest_rule.json's disagreement-resolution rule:

   python3 -m lexeme_aligner.merge_align --iso swe --methods eflomal,gloss,gapfill \
     --contest-rule contest_rule.json      # → align_merged_swe_*.jsonl, then export --methods merged
   

   Labelled lossy on purpose — it drops valid alternatives; use it only when you want exactly one row. contest_rule.json is published here for transparency (see below) but its keys need per-occurrence data (eflomal's raw score, gloss's match-type) that this dataset's aggregated rows don't carry — you can't run this rule yourself on the Parquet alone, only approximate its spirit via hi_conf/count.

Companion reference resources (root of this dataset, small + committed)

Four small JSON files sit alongside manifest.json — each is knowledge our own pipeline uses internally that can't be derived from the row data itself (unlike strong/share above), so it's published outright rather than left for every consumer to rediscover independently.

file	keyed by	directly usable on this dataset's rows?
light_lexemes.json	lexeme	yes
hebrew_lexeme_strong.json	lexeme	yes
greek_morph_strong.json	lexeme + source grammar code	no — needs external morphology
contest_rule.json	eflomal score + gloss match-type	no — needs per-occurrence data

• light_lexemes.json — {lexeme: avg_target_dominance}, ~545 entries. Source lexemes so semantically broad (light verbs like Hebrew הָיָה/"to be", Greek γίνομαι/"to become"; generic nouns) that no single target rendering dominates in any language — a single-method row for one of these is weaker evidence than for an ordinary content word. Directly applicable: if row.lexeme is a key here, prefer rows that are cross-method-corroborated (see above) over trusting a lone gloss row. Computed by cross_lang_prior.py from cross-lingual target-dominance across every aligned language.
• hebrew_lexeme_strong.json — {lexeme: strong}, a small, verified override for the handful of Hebrew lexemes where the mechanical strong derivation (above) would be wrong: our spine's own "equivalence-canonicalization" occasionally rolls two genuinely distinct lexemes onto one bare Strong's number and doesn't always pick the number in wider use (verified case: hbo:4714 "Egypt" mechanically derives to H4713 "Egyptian" — wrong; this table corrects it to H4714, matching Clear-Bible gold's own usage 1,633:55). Directly applicable: check this table BEFORE the mechanical derivation; only 1 entry currently, scoped to verified cases, not a blanket table (a broad, unscoped sweep for this was tried and produced nonsense — see hebrew_lexeme_strong.py's docstring).
• greek_morph_strong.json — {"lemma_strong|grammar_code": traditional_strong}. Clear-Bible's gold uses the traditional Strong's Concordance numbering for irregular Greek verbs — separate numbers per tense/person (εἰμί: G2258 imperfect, G1526 present-3pl, etc.) — while lexeme's mechanical rollup collapses them all to one lemma number (G1510). This table recovers the traditional number if you have the source occurrence's own morphological parse (e.g. V-IIA-3S) in the same coding convention as globalbibletools/data's hbo+grc source (see greek_morph_strong.py) — this dataset's rows don't carry that, so it's not self-contained, but it's the exact table our own benchmark scoring uses, published for anyone doing source-morphology-aware work.
• contest_rule.json — {"score <eflomal_score> | <gloss_match_type>": "ef"|"gl"}. An empirically validated (leave-one-out tested across 10 gold languages), universal rule for which method's answer to trust when eflomal and gloss disagree on the same source token. Published for transparency about what the "merged best-pick" derived view (above) actually does — not directly runnable against this dataset's aggregated rows (see the table above), since both tier keys need per-occurrence values this dataset doesn't carry.

Layout — why the bulk data isn't in git

lexeme-alignments/
  README.md                    # committed — this file
  manifest.json                 # committed — per-language metadata + content hash (the durable record)
  light_lexemes.json            # committed — see "Companion reference resources"
  hebrew_lexeme_strong.json     # committed — see "Companion reference resources"
  greek_morph_strong.json       # committed — see "Companion reference resources"
  contest_rule.json             # committed — see "Companion reference resources"
  iso=<iso>/                   # GIT-IGNORED — bulk data, published out-of-band (HF / object storage)
    data.parquet

manifest.json is git's small, diffable record of what exists and what it hashes to; each partition is keyed by its content_sha256.

Provenance & quality

Per-language provenance (methods present, per-method row counts, testament, counts, hi_conf_ge_0.9, spine tags, content hash) lives in manifest.json. Every language is produced by the same pipeline, validated against Clear-Bible manual gold where it exists — token-weighted top-1 of ~92–97% (Strong's grain) / ~89–92% (lexeme grain — the anchor's headline; docs/benchmark.md). Languages without usable gold (ind; rus, whose only manual reference is itself mis-aligned) run the identical pipeline and are not lower quality — simply un-cross-checked. We do not stamp a verified/unverified tier. Your confidence signal is the same for every language: the row-level method / hi_conf / count (plus the derived share — see above). These are raw aligned counts, not hand-checked.

Reproducibility (content-addressed)

The statistical aligner (eflomal) seeds from /dev/urandom, so regeneration varies ~1% run-to-run. This is a content-addressed release: inputs are pinned (spine tags + each source text's sha256, data/pins/), and each partition is fixed by its content_sha256 in manifest.json — that hash is the identity of what was released. Consume a specific release by its hash; a rebuild won't match byte-for-byte.

Authentication & publishing (one-time)

python3 -c "from huggingface_hub import login; login()"        # cached → ~/.cache/huggingface/token
python3 -m lexeme_aligner.export_lex --iso <iso> --lang-name <Name> \
  --publish bcv-commons/lexeme-alignments --create

Use a fine-grained write token scoped to the target dataset. The push uploads this language's partition + the shared manifest.json/README.md/companion resource files (light_lexemes.json, hebrew_lexeme_strong.json, greek_morph_strong.json, contest_rule.json — global, not per-language, but small enough to just re-upload each time so they never drift out of sync); other languages' partitions are untouched.

License

This catalogue is CC0-1.0. It is derived, factual data — lexeme ids, Strong's rollups, alignment counts, share/hi_conf statistics, method tags, and a de-arranged type-level list of word forms. It does not reproduce the running text of any translation (no verse refs, no word order), so the copyrightable expression of the sources is not present.

Each surface is nonetheless a word form from a source translation, and those keep their own licenses. Every language's manifest.json entry carries a source pointer (provider/edition/license_url) — follow it for the authoritative terms. Pointing to a source does not by itself grant permission to derive from it; for any source whose terms restrict derivatives, obtain that separately.