bcv-commons/hebrew-word-senses

Hebrew Word Senses

Which Biblical Hebrew words have more than one sense, and which sense each occurrence belongs to: 9,355 MACULA lexemes, 1,875 of them split into several senses, 311,297 occurrences assigned. CC BY 4.0. No BHSA content.

What changed (2026-10-09)

Senses of verbs are now per stem. A verb that occurs in two or more stems (binyanim: Qal, Piel, Hiphil ...) is split per (lexeme, stem), from the same kind of evidence as before (which occurrences translators render alike), now also using the Old Testament editions of 60 languages' per-occurrence alignments from lexeme-aligner. 908 verb lexemes are re-split into 3,872 stem-bound senses; senses.tsv has a new stem column for them (empty for every other lexeme). Sense numbers of those 908 verbs changed (a lexeme's senses are numbered by size within each stem, stems in order of size). All other lexemes keep their senses. 102 occurrences of these verbs that had a sense before, and whose stem is not given in the source, now get none. Against the same references: UBS ARI 0.316 (was 0.317), Mahberet Menahem ARI 0.208 (was 0.170); pair F1 is a little lower (0.596 / 0.525 against 0.613 / 0.577) because the stem split adds divisions the references do not make.

Labels. A label wrote the most common English surface of its rendering across the whole Old Testament instead of the one its own occurrences carry, and the English normalisation merges words that share their first five letters. So the verb קדשׁ (to be holy) had a sense labelled "sanctuary", and the names Abishur and Abishua were labelled "abishai". Labels now show the surface of the sense's own occurrences ("sanctified"); 2,794 of the earlier 12,088 labels changed. Reported by the lexeme-aligner project.

What changed (2026-10-05)

Senses now draw on the per-occurrence alignments of 210 languages' Old Testament editions (the lexeme-aligner's), not only ten manually aligned translations: agreement with the UBS dictionary's senses rose from 0.206 to 0.317 and with Mahberet Menahem's from 0.136 to 0.170. Labels avoid giving two senses of one word the same name.

Earlier (2026-10-04)

The previous release (2026-08) was keyed on ETCBC BHSA lexeme ids and clustered on BHSA clauses, so it was bound by BHSA's CC BY-NC-SA licence (relabelled accordingly, now replaced). This release is built from openly licensed data only, keyed on MACULA lexemes. It agrees better with the UBS dictionary's senses and as well with Mahberet Menahem's (below). For expert senses keyed on the UBS Dictionary of Biblical Hebrew, see bcv-commons/senses-attested (CC BY-SA 4.0).

Method

Occurrences of a word that translators render alike share a sense. For רוּחַ, translators in over two hundred languages write "spirit / espíritu / esprit / дух" in some verses and "wind / viento / vent / ветер" in others; the renderings of one sense co-occur on the same Hebrew word across languages, those of different senses do not. Per lexeme: a graph of (language, rendering) features from per-occurrence alignments, edges weighted by how often two renderings occur on the same Hebrew token; communities of that graph are the senses; each occurrence goes to the sense its renderings vote for. Grammatical number and pronouns are normalised away, so plural forms and suffixed forms ("his hand") do not form senses of their own. The main sense is labelled by its most frequent English rendering, each other sense by its most distinctive one (numbered, "needy (3)", when English uses the same word for both). Names keep their splits: translations often spell same-named people or places differently (Abimelech of Gerar, Abimelech son of Gideon). For the stem-split verbs a sense is labelled by the frequent rendering most specific to it against the verb's other stems (a passive stem's participle reads "be gathered"); the label of one stem therefore differs from another's only where their translations do.

Examples: רוּחַ spirit / wind; דָּבָר word / thing / matter; אֶרֶץ earth / land; נֶפֶשׁ soul / life / person.

Quality

Measured against two independent expert sense divisions, adjusted Rand index (0 = no better than treating each word as one sense):

reference	this release	previous release (BHSA)	MACULA English gloss
UBS Dictionary of Biblical Hebrew per-token senses (630 polysemous words)	0.316	0.161	0.158
Mahberet Menahem's sense divisions (10th c., 231 words)	0.208	0.154	0.071

Occurrences whose translations give no content word (about a quarter, mostly where a word is rendered only by grammar) get no sense. Splits are conservative where translations do not separate senses; verb "senses" can partly follow tense or stem rather than meaning. Treat sense as a translation-based grouping, not a dictionary sense.

Files

senses.tsv: one row per (lexeme, sense)

column	meaning
lexeme	MACULA lexeme id (hbo:7307; homographs carry a letter suffix)
strong / lemma	Strong's number and pointed lemma
sense	sense number within the lexeme; 1 = the most frequent
label	the sense's most distinctive English rendering
count / share	occurrences in this sense, and their share of the lexeme's aligned occurrences
stem	for the senses of verbs that occur in several stems: the stem (qal, piel, hiphil ...) this sense belongs to; empty otherwise

occurrences.parquet: one row per aligned occurrence: key (MACULA word-part id, as in macula-hebrew), book, chapter, verse, lexeme, sense.

Sources and licence

CC BY 4.0. Built from per-occurrence alignments by Clear-Bible/Alignments (CC BY 4.0; ten languages: English BSB and YLT, Spanish RV09, French LSG, Portuguese, Russian Synodal, Arabic Van Dyke, Hindi, Bengali and Assamese IRV, Hausa), per-word glosses from Global Bible Tools (CC0; Spanish, Portuguese, Hindi, Hungarian, Nyamwanga and others) and MACULA Hebrew lexemes (CC BY 4.0), and the per-occurrence alignments of 210 languages' Old Testament editions by bcv-commons/lexeme-aligner (CC0), read together with those editions' text locally. No translation text is included beyond the one-word English labels (from the public-domain BSB and YLT).

Provenance

Built in bcv-commons/bcv-query: shoresh/macula/build_compact_renderings.py --languages 999, shoresh/macula/build_rendering_senses.py --gbt --compact macula/data/compact_renderings_999.pkl, shoresh/macula/export_hebrew_word_senses.py.