bcv-commons/semantic-neighbors

Hebrew Semantic Neighbor Pairs

6,671 Strong's-number pairs of Biblical Hebrew words judged to be synonyms or to share a tight semantic concept, each passing at least one of three independent confidence gates. CC0, built only from public- domain, CC0 and CC BY sources plus model judgments (list below).

What changed in this release (2026-10)

• No BHSA input. The ETCBC BHSA database is licensed for non-commercial use, which does not fit a CC0 dataset. Earlier releases used it for clause contexts and for syntactic coordination/apposition pairs. This release embeds MACULA's Westminster Leningrad Codex text (CC BY 4.0) in word-centred windows instead, and drops the syntactic-pair signal.
• No Wiktionary input. The cross-lingual/Wiktionary-root corroboration family (CC BY-SA source) is not counted toward any gate; 78 pairs that depended on it were dropped.
• Homograph routing. Evidence keyed by Strong's number (BDB roots, LLM judgments) now attaches only to the homograph it concerns, using MACULA's lexeme distinctions (e.g. H7462 "shepherd" vs "associate with" no longer share evidence).
• No SDBH figures. Quality is no longer reported as agreement with the Semantic Dictionary of Biblical Hebrew; see "Quality" below.

Method

The pairs are the confident subset of a larger semantic-neighbor graph combining these signals: distributional embeddings (BEREL 3.0 over Hebrew word windows), LXX co-rendering, gloss overlap, an LLM scholarly-prior pass, Brown-Driver-Briggs (1906) etymological roots, T'OMIM poetic-parallelism pairs, and candidates from Radak's Sefer HaShorashim. A pair is published if it passes one of three gates:

• cross_signal: asserted by at least two methodologically independent signal families.
• llm_verified: a single-family pair individually judged by an LLM ("are these genuinely synonyms in Biblical Hebrew usage", strict); only "yes" verdicts, and only pairs the graph still links.
• sefer_hashorashim_verified: words Radak discusses together within one root entry, LLM-verified the same way.

1,411 pairs pass more than one gate (gate column, "+"-joined); treat those as the highest-confidence layer.

Columns

column	meaning
strong_a / strong_b	Hebrew Strong's numbers (H####)
gate	which check(s) the pair passed: cross_signal, llm_verified, sefer_hashorashim_verified, or a "+"-joined combination
n_families	how many independent signal families asserted the pair (0 if only LLM-verified)
llm_verdict	yes if LLM-verified (either LLM gate); blank if published through cross-signal agreement alone

Quality

Measured on the Hebrew text itself rather than against a human taxonomy: on books held out from the build, published pairs share verb-argument slots (same verb, same role) more often than frequency-matched random pairs (mean slot-profile similarity 0.028 vs 0.012; held-out slot co-occurrence 5.4% vs 5.0%, a weaker separation). That check uses syntactic annotation for measurement only; it plays no part in building the dataset.

The groups built from the same graph were also tested for usefulness to readers: labels for a marked word in a verse were picked correctly 83–88% of the time by two independent LLM judges (chance 25%), and a planted outsider was found in a group 79–85% of the time (chance about 20%).

These are internal checks, not an audit; a human check is in progress.

Sources and licences

MACULA Hebrew (WLC text, lexemes, glosses; CC BY 4.0), BEREL 3.0 (Apache 2.0), LXX bridge (CC BY 4.0), Brown-Driver-Briggs (1906, public domain, via OpenScriptures CC BY 4.0), T'OMIM (CC BY 4.0), Sefer HaShorashim (Radak, Public Domain, via Sefaria), and LLM judgments (Claude). No BHSA, SDBH, or Wiktionary data is an input.

Provenance

Built in bcv-commons/bcv-query: shoresh/macula/build_bhsa_free_contexts.py, build_semantic_neighbors.py --macula-contexts --no-structural --parallelism-tomim-only --no-xling --route-homographs, build_confidence_tiers.py, build_published_pairs.py --llm-needs-tier-support --exclude-families corroborated.