bcv-commons/hebrew-phrase-types-model

Hebrew Phrase Types: a model layer

Model output, not a fact layer. This dataset gives semantic types for the 50,829 clause-level prepositional phrases of the Hebrew Bible: the phrases MACULA Hebrew gives the role pp.

• Two language models read them. Jev (TypeSafe) read every phrase; Claude Opus 5.5 read, as a second judge, the 24,057 phrases that a fixed rule calls unsure.
• No person has reviewed the rows.
• It is published apart from the bcv-commons fact layers (the clause layouts, passage facts, word parts) and is never merged into them.

Version	v0.1 (schema 1); from make-tanakh-research's table version 2.1 (sha256 53aa12e53c98b335d856c1fe4777e00639cf494b57c37fd4cf60f3b2cffaeda5)
Built	2026-11-13 by shoresh/macula/release_phrase_types_model.py (bcv-query df4cdd3)
Rows	50,829 phrases: 44,291 with a nominal head, 6,538 preposition + infinitive or participle
Licence	CC BY 4.0

Why

MACULA Greek types its adjuncts; MACULA Hebrew gives a clause-level prepositional phrase the role pp and no type. We found no openly licensed source that types them. This layer fills that gap as a judgement, with its confidence stated, for readers and drafting tools that need to know whether "to Jerusalem" is a goal and "to Abraham" a recipient.

The labels

Fifteen labels:

• Recipient, Beneficiary, Goal, Place, Source, Time, Instrument, Accompaniment, Manner, Comparison, Relation, Purpose, Possessor;
• Object (the verb's prepositional object);
• Other.

Each label belongs to one of four groups:

group	labels
participant	Recipient, Beneficiary, Object, Possessor, Accompaniment
space	Goal, Place, Source
time	Time
circumstance	the other six

The file

data/hebrew_phrase_types.parquet, one row per phrase:

column	meaning
phrase_key	the phrase: <first word>-<last word>:pp:<level>, MACULA Hebrew word ids. level counts from the top within a chain of tree nodes over the same words; it is 0 for nearly all.
first_word	the phrase's first word id (12 digits, MACULA Hebrew)
preposition	the preposition's lemma
head	nominal, or verb (preposition + infinitive or participle)
label	the published label: Opus's where Opus was asked, else Jev's
basis	how label came about (below)
group	the coarse group of label
jev_label, jev_confidence	Jev's choice and its probability
jev_second, jev_second_p	Jev's second-likeliest label and its probability
jev_group_p	the probability Jev gave to the group of label, summed over its labels
opus_label	Opus's label where Opus was asked, else empty

jev_confidence describes jev_label. Where basis is opus, it does not describe label.

Use basis

basis	rows	share	what it means
jev	26,772	53 %	Jev alone, confident; right in about 96 % of a random sample
jev+opus	12,643	25 %	asked of both, and the two models agree; right in 97–98 % where the reference decides
opus	11,414	22 %	asked of both, and they differ; Opus's label stands. These are the hard phrases: one in six is contested even between two careful labellings, and the label there is one of two defensible readings.

How good it is

The yardstick is a simple random sample of 450 phrases (300 with a nominal head, 150 with a verb head).

• Each phrase was labelled twice by Claude Opus 5.5, in two orders. Where the two labellings agree (283 and 133 phrases), that label is the reference; on 17 + 17 phrases they differ, and those are "contested".
• Caution: the second judge is the same model as the reference's labellers. Where label is Opus's, "equals the reference" measures Opus's consistency with itself, not an independent check. The independent evidence is where two different models agree, and the comparison with a human analysis below.

sample	n	equals the reference	contested	differs	share, of those with a reference
nominal heads, label	300	273	17	10	0.965
nominal heads, Jev alone	300	242	17	41	0.855
verb heads, label	150	130	17	3	0.977
verb heads, Jev alone	150	87	17	46	0.654

Overall. Weighted by the two head kinds, label equals the reference in 96.7 % of the phrases where the reference decides (Jev alone 82.9 %). Counting contested phrases as unconfirmed, it is 90.4 % of all phrases (Jev alone 77.8 %). The group of label equals the reference's group in 98.2 % (nominal heads) and 98.5 % (verb heads).

Jev's confidence is calibrated where Jev stands alone. Share of Jev's labels that equal the reference, by confidence:

jev_confidence	nominal heads	verb heads
below 0.5	0.44	0.25
0.5–0.7	0.84	0.73
0.7–0.9	0.95	0.83
above 0.9	0.98	0.97

Against a human analysis. This compares nominal heads with the ETCBC's BHSA phrase functions. It gives figures only: no BHSA value is in this dataset.

line	label	Jev alone
their Time phrases are typed Time	0.983	0.977
their Location phrases are typed Place	0.884	0.838
their Location phrases are typed Place, Goal or Source	0.986	0.968
typed Recipient or Object is their Complement	0.941	0.896
typed Time is their Time	0.791	0.804

Where the two models differ, Opus's label is nearer the human analysis on every line. "Typed Time is their Time" is low on the unsure phrases for both models: these are phrases BHSA files as adjunct or complement ("to your generations", "on this day").

A registered check fired, and we say so. Before the second pass we wrote down a rule: the merged layer's Time and Place figures must not fall more than 0.05 below Jev's figures at confidence ≥ 0.7.

• Two of the three figures held. "Typed Time is their Time" fell by 0.070. The cause is that the check compared the whole layer with Jev's confident part, so it fell when the unsure phrases came in, whoever labels them: Jev alone on all phrases gives 0.804.
• On like phrases the second judge is the better one. The merged label is published; jev_label and opus_label are both given, so that either reading can be used.

Known limits

• The fifteen labels are our list. The judges differ most at these boundaries:
  • Recipient / Beneficiary;
  • Goal / Place;
  • Purpose / Manner (for לֵאמֹר and other infinitives);
  • Object / Beneficiary;
  • Recipient / Object.
• Recipient against Beneficiary is a convention, not an error. For "bear to", "forgive" and "give", the layer says Recipient throughout.
• Only clause-level pp. Phrases inside noun phrases, predicates of verbless clauses and members of coordinated phrases are not typed here.
• Complement against adjunct is not in this layer.
• No person has reviewed the rows. Corrections are welcome as issues on bcv-commons/bcv-query, most of all on the opus rows.

Sources and licence

Inputs sent to the models	the Hebrew text's words with MACULA Hebrew's lemmas and glosses (CC BY 4.0, © Clear Bible, Inc)
Phrase boundaries and word ids	MACULA Hebrew (WLC lowfat), CC BY 4.0
Labels	Jev (TypeSafe) and Claude Opus 5.5, prepared by make-tanakh-research for bcv-commons
Witness for the figures only	ETCBC BHSA (CC BY-NC 4.0); none of its values are in the data

This dataset is licensed CC BY 4.0. Please credit "Hebrew Phrase Types (model layer), bcv-commons (bcv-commons/hebrew-phrase-types-model v0.1)" and MACULA Hebrew, © Clear Bible, Inc.

Versions

• v0.1 (2026-11-13): first release (research table version 2.1).
• schema in build.json is the version of the file shape. It is raised only when a column is renamed, dropped or changes meaning.