bcv-commons/senses-attested

senses_attested — attested target renderings per lexeme sense

The empirical evidence layer produced for shoresh (bcv-query data-contract): for a lexeme in a disambiguated (binyan, sense), which target-language words attest it, with counts. It is the supply that fills shoresh's senses_i18n/_gaps demand and cross-checks the llm_strongs_glosses predictions — it does not replace shoresh's curated senses_i18n/<iso>.tsv; consumed as an HF Parquet dataset.

Schema (per row)

column	meaning
lexeme	MACULA lexeme (the anchor), e.g. hbo:0006
stem	MACULA binyan (qal/piel/hiphil/…); empty for non-verbs
sense	sense number (ordinal) — see licensing
surface	attested target rendering (lowercased)
count	times this (lexeme, stem, sense) → surface was aligned
share	count / Σ count for that (lexeme, stem, sense) within one base_text
method	alignment method (eflomal)
source_corpus	the original Hebrew corpus (e.g. WLC)
base_text	the target edition attested (e.g. ind_C01) — the per-row provenance dimension

Key: (lexeme, stem, sense) — MACULA lexeme (anchor; BHSA lex dropped) + MACULA binyan + sense number, read inline from the enriched lexeme-spine.db. OT/Hebrew only (senses are Hebrew; Greek tokens carry none).

Multi-version: base_text is per-row, so several translations of a language are pooled into one iso=<lang> partition — a union of per-edition runs, each row tagged by edition; share stays per-edition. Cross-edition agreement (how many base_texts attest a given (lexeme,stem,sense)→ surface) is the confidence signal, derivable directly from the rows. (Swedish iso=swe pools swe_fol Folkbibeln + swe_svk Kärnbibeln.)

Removal / takedown policy

Each row is a per-edition attestation carrying its base_text, so a rights-holder can request removal and it is a clean row-drop + republish (the dataset is content-addressed via each partition's content_sha256). Because most (lexeme,stem,sense)→surface facts are attested by more than one edition, dropping one edition typically leaves the linguistic fact intact via the others — properly attributed. Rows are never re-emitted with provenance stripped: a removed attestation is removed, not anonymized.

Removals are driven by a committed, auditable config: data/senses_exclude.json (read automatically on every build). A row is dropped if it matches any rule; a rule matches when all its stated fields equal the row's — fields lexeme, stem, sense, surface, base_text, omit to wildcard:

{"exclude": [
  {"base_text": "swe_fol"},                       // drop a whole edition
  {"base_text": "swe_fol", "surface": "herren"}   // drop one surface within an edition
]}

After exclusion, survivor shares renormalise (per edition), so a removed row leaves no residue; the manifest records excluded: {rules, rows_dropped} for the audit trail. To action a takedown: add a rule, re-run senses_attested for the affected language, republish.

Licensing — CC-BY-4.0, deliberately label-free

The key is MACULA-derived (lexeme + binyan), so this dataset is CC-BY-4.0 — attribute Clear-Bible MACULA. We carry the sense number only and no English sense label: shoresh's sense labels are UBS-MARBLE "used with permission" (not redistributable), so the payload is pure attestation (lexeme, stem, sense#, surface, count) — CC-BY clean. Regenerate: python -m lexeme_aligner.senses_attested --iso <iso> --method eflomal. Same git-ignored-Parquet + committed-manifest.json layout as lexeme-alignments.