bcv-commons/hebrew-nt-analysis

Hebrew New Testament Analysis

Hebrew translations of the New Testament, analysed word by word as MACULA Hebrew analyses the Old Testament. Each word is split into its parts (prefixes and suffixes are parts of their own), and each part has a class, a lemma, a MACULA lexeme id and a Strong's number. Each verse is also divided into clause segments, with their depth and parent inside the verse. It is meant for aligning a Hebrew New Testament with the Greek, for lexical study across the two Testaments, and as a Hebrew reading of the New Testament's clause structure.

Release	v0.1, built 2026-10-23 from bcv-query f3101cf
Editions	heb_scg: Salkinson–Ginsburg (1885/86)
Size	115,262 words, 186,994 word parts, 35,475 clause segments

Most of this is model output. The New Testament has no MACULA analysis. Words whose form occurs in MACULA's Old Testament with one near-unanimous analysis take MACULA's own (source = "macula", 92,983 words, 80.7%). The rest, 22,279 words, and every clause segment come from a tagger trained on MACULA Hebrew (source = "model"). It has not yet been checked by hand on the New Testament: see Quality.

Files

File	Rows
heb_scg/words.parquet	one per word part
heb_scg/clauses.parquet	one per clause segment
heb_scg/manifest.json	per book: the text fingerprint, verse and word counts; the model's SHA-256; tokenizer and verse-text versions
SHA256SUMS	SHA-256 of every file

The words

A word is what the bcv-commons tokenizer (v2, bibles verse-text 1.0.0) gives: a run of letters, split at spaces, maqaf and punctuation. The verse text and the word positions are the same as in bcv-commons/lexeme-aligner's alignments of the edition: i is the aligner's token number, and the text fingerprints in manifest.json match the aligner's file hashes.

Column
book, chapter, verse	USFM book code; the edition's chapter and verse
i	the word's position in the verse (0-based, the aligner's token)
part	the part's position in the word (0-based)
token	the word as the tokenizer gives it: unpointed
word	the word with its vowel points
text	the part with its points. Concatenating the parts gives word. The article after a preposition is a part without letters (""), as in MACULA: בַּמָּשִׁיחַ is ב + the article + מָּשִׁיחַ
class	MACULA word class: noun, verb, prep, cj, art, pron, adj, adv, num, rel, ptcl, ij, om (object marker), x; verse-number for the verse number some editions carry at the start of the verse text (Salkinson: "א", "ב", … "טו"), which is not part of the text
lemma	MACULA's lemma. Where no lemma is given (lexeme empty), the part's own letters, unpointed
lexeme	MACULA lexeme id (hbo:0559); empty for a name or word MACULA does not have, and where no known lemma fits the letters
strong	Strong's number as MACULA gives it (0559, 3807a)
role	MACULA role, where the part heads one (v, s, o, p, adv, …)
source	macula: the analysis of this form in MACULA's Old Testament (at least 90% of its occurrences analysed alike); model: the tagger's
p_split	for model words, on part 0: the tagger's confidence in the split (among the splits that fit the word)
form, person	on the verb part: MACULA's verb form (qatal, wayyiqtol, yiqtol, weqatal, imperative, jussive, cohortative, participle active/passive, infinitive construct/absolute) and the person (first, second, third)
seg	the clause segment of the verse the word stands in (1-based; see clauses.parquet)

The clauses

A clause segment is a run of whole words inside a verse that belongs to one clause, as in the Hebrew Bible Clause Layout built on MACULA's trees; a clause interrupted by another gives two segments.

Column
book, chapter, verse, seg	the segment (1-based in the verse)
first_i, last_i	its first and last word (i)
depth	clause ancestors inside the verse, as MACULA's tree counts them (0 = on the verse's own group)
parent	where its parent clause stands, in segments from this one: -1 the one before, +1 the one after, -4+, +2+; none on the verse's own group
kind	VC (with a verb) or NC
rule	the MACULA clause rule, the order of its constituents (V-S-PP; the 60 most common rules, others other)
att	the MACULA attachment kind where the segment meets an earlier segment of the verse in the tree ("parent rule | child index | class | role | opening word", as the clause layout's [tree] setting reads it; - none; the 300 most common, others other)

How it was made

The tagger (python -m macula.clause_tagger in bcv-query) reads a verse with a Hebrew BERT (dicta-il/dictabert, CC BY 4.0), whose tokenizer drops the vowel points, and with a character channel that reads each word with its points. It was trained on MACULA Hebrew's Old Testament (commit 47db250): the words, their parts, classes, lemmas and roles, MACULA's verb forms, and the clause segments, depth, parents, clause rules and attachment kinds of MACULA's trees as the clause layout reads them. No BHSA data, and no English, went in. A lemma has to fit the part: among the model's best lemmas the one whose strong consonants (not א ה ו י נ) the part carries most fully; where none explains a part of three or more such consonants by more than one, no lemma is given. The model file (SHA-256 6a5877b528523f91e4b2d3ad3ed84871062293e1e782911c3a6adaf76fbef7cf) is not part of this release.

Quality

Measured on 856 Old Testament verses the tagger was not trained on (Gen 37–45, Ruth, Jonah, 1 Sam 1–10, Deut 5–11; narrative and discourse, the closest Old Testament match to a New Testament in this style), against MACULA, with the same lookup of known forms as here (from the training books only):

	All words	Forms not in the training books (12%)
Split into parts right	97.9%	82.7%
Class (rightly split words)	99.0%	95.0%
Lemma (rightly split words)	95.3%	66.3%
Lexeme given / right when given		86.7% of parts / 76.5%

Clauses
Segment boundaries between words (F1)	94.9%
Segments reproduced exactly	88.6%
Depth in the verse, on exactly matched segments	84.4%
Parent in the verse, where there is one (664 segments)	89.3%
Verb found / verb form / person	99.1% (recall) / 96.0% / 97.6%

For comparison, a modern-Hebrew parser (Parsan, DictaBERT-based) on the same verses: segment boundaries 77.1%, segments exact 56.3%, verb form 92.5%.

The forms not in the training books are the closest stand-in for the New Testament's own words (names, New Testament vocabulary, new prefixed forms); on them expect the lower column. The New Testament has not been scored, since it has no reference analysis. To favour precision, filter on source, p_split and an empty lexeme.

Sources and licence

Part	Source	Licence
Salkinson–Ginsburg Hebrew New Testament	via bible.helloao.org (heb_scg), from eBible.org (hebsg)	public domain
Word parts, classes, lemmas, lexemes, verb forms, clauses (training data and the forms looked up)	MACULA Hebrew (Clear Bible / Biblica), WLC/lowfat, commit 47db250	CC BY 4.0
Base model	dicta-il/dictabert	CC BY 4.0
Verse text and words	bcv-commons/bibles verse-text 1.0.0	CC0 1.0
This analysis	bcv-commons	CC BY 4.0
The code (bcv-query)	bcv-commons	MIT

Attribution when you use this dataset:

Hebrew New Testament Analysis (v0.1), bcv-commons, CC BY 4.0. Analysis trained on MACULA Hebrew, Clear Bible / Biblica (CC BY 4.0). Text: Salkinson–Ginsburg Hebrew New Testament (public domain).

Feedback

Issues and requests: bcv-commons/bcv-query. Please name the edition, the verse and the word, and what you expected.