source-ontology / README.md
larsgson's picture
source-ontology v0.1 (code 341e013)
613b05b verified
|
Raw History Blame Contribute Delete
8.81 kB
metadata
license: cc-by-sa-4.0
language:
  - hbo
  - grc
pretty_name: Source ontology
tags:
  - bible
  - hebrew
  - greek
  - septuagint
  - ontology
  - knowledge-graph

source-ontology v0.1

One id space for the meaning side of the Tanakh, the Septuagint and the Greek NT: lexemes, UBS senses and domains, concepts, the words (mentions) that evoke them, the persons, places and things the words refer to, how lexemes are rendered across the languages, which words are aligned to which, and one proposition per clause with its predicate, arguments and links. The tables join bcv-commons datasets on shared keys (MACULA word ids, Strong's numbers, UBS codes); nothing here is new judgment except what those datasets already hold.

No English. Lemmas are Hebrew and Greek (MACULA's homonym markers "(I)", "(II)" kept); senses and domains are UBS codes (their English glosses and definitions are never read); feature values are the source datasets' codes; connectives are written in Hebrew or Greek.

Ids

id for
lex:H0001, lex:G3962 a lexeme by Strong's number (Aramaic under H)
con:H1.1 a concept (hebrew-word-concepts-model / greek-nt-concepts-model); con:H0.* belong to words without a Strong's number
sense:<UBS sense id> a UBS dictionary sense
dom:sdbh:<code>, dom:ln:<code> a UBS domain: SDBH lexical domain (Hebrew), Louw-Nida domain or subdomain (Greek)
w:010010010011 a Tanakh word (MACULA Hebrew id)
w:40001001001 a Greek NT word (MACULA Greek id, Nestle 1904, without its leading n)
w:glx:104873137 a Septuagint word (GLAUx id)
w:heb_scg:MAT.1.1.3.0 a word part in a Hebrew NT (edition, book, chapter, verse, word, part)
ent:e0123456789 an entity (person, place, group, deity, realia, flora, fauna), the codes of bcv-commons/entity-mentions
prop:<clause key> a clause as a proposition; the key is the clause's first MACULA word, as in every bcv-commons clause dataset

Tables (data/)

file rows
classes.parquet lexeme 13,786, sense 18,606, domain 1,125, concept 8,555 id, class, lang, lemma
relations.parquet see below subject, relation, object, kind (asserted / derived), method (source / computed / model), source (dataset@version), confidence, count, share, tier
mentions.parquet Tanakh 469,476, Greek NT 137,779, Septuagint 491,860 mention, text, lexeme, concept with concept_labeller and concept_confidence, sense
aligned.parquet Septuagint 345,461, Salkinson 123,728, Delitzsch 64,531 mention, other, edition, basis (the source's own label), method
propositions.parquet Hebrew 87,582, Greek NT 29,682 prop, lang, ref and the clause's features (below)
prop_relations.parquet see below subject, relation, object, tier, gram, word, source, target_basis
prop_accuracy.parquet one per checked feature lang, feature, figure, n, sample, basis
entities.parquet 5,243 id (ent:e…), class entity, type, gender, lemmas
entity_relations.parquet refers_to 138,428, kinship 10,925 subject, relation, object, kind, route, agree

Relations:

relation method rows
in_domain source 33,124
sense_of source 18,605
from_sense source 8,555
concept_of model 8,519
rendered_by computed 6,438
rendered_by model 4,906
broader source 1,028

concept_of links a concept to its lexeme, sense_of a sense to its lexeme, from_sense a concept to the UBS sense it was made from, in_domain a sense to its domain, broader a domain to its parent, rendered_by a lexeme (or concept) to a lexeme of the other language with the tier, count and share of the bridge it comes from.

Model rows. Rows with method model come from bcv-commons datasets made by a language model (the -model datasets named in source: concepts, concept labels per word, model-chosen lexeme pairs). They were not checked by a person; see each dataset's card for the model and its figures. Filter on method to leave them out.

Entities

From bcv-commons/entity-mentions v0.1 (code 341e013; ACAI with MACULA's reference chain). refers_to links a mention to its entity, with the mention's kind (explicit, pronominal, subject, other), the route that found it (acai, chain, acai+chain) and whether the routes agree (same, differs, acai_only, chain_only). Where the routes disagree, both rows are here. kin:<relation> links an entity to another (ACAI's kinship codes: father, mother, partner, offspring, sibling, tribe, group_origin, birth_place, death_place, possibly_same_as).

Propositions

One row per clause of the Hebrew and Greek NT clause layouts. Features (codes, from the dataset in build-propositions.json): voice, neg, neg_scope, force, kind, interrog, speech (speech level), speaker_id, line (narrative line), genre, front, focus, subj_switch, part_switch, causative (clause-columns v1.2); time, aspect, time_basis (time-aspect v0.2).

relation source rows
has_argument argument-roles 156,658
linked_to clause-connectives 117,198
has_predicate clause-columns 95,762

has_predicate points at the main verb's mention; has_argument at an argument's head word, with its role (actor, undergoer, location, recipient …) in tier and its grammatical relation in gram (argument-roles v0.2); linked_to at the clause it attaches to in the clause layout (target_basis layout), else at the clause before it in the same book (target_basis previous), with the link type in tier, the relation in gram and the connective as written in word (clause-connectives v0.1).

How far the features can be trusted

prop_accuracy.parquet holds the figures of an audit of the clause columns (S2): blind samples of 100 clauses per language, judged from the Hebrew or Greek text alone, exact agreement with the dataset. sample names the sample (seed and the clause-columns version it checked; the Hebrew columns are unchanged in v1.2). basis: same_word (the judge's main verb is the same word), verb_morphology (follows the main verb's MACULA morphology), present (a negator is there or not), same_scope (predicate or constituent negation), both_named (subject switch, where judge and dataset both name a subject).

lang feature figure n sample basis
hbo has_predicate 0.970 100 s2:1119:v1.1 same_word
hbo voice 0.970 100 s2:1119:v1.1 verb_morphology
hbo neg 1.000 100 s2:1119:v1.1 present
hbo neg_scope 0.980 100 s2:1119:v1.1 same_scope
hbo subj_switch 0.931 87 s2:1119:v1.1 both_named
grc has_predicate 1.000 100 s2:2611:v1.2 same_word
grc voice 1.000 100 s2:2611:v1.2 verb_morphology
grc neg 1.000 100 s2:2611:v1.2 present
grc neg_scope 0.990 100 s2:2611:v1.2 same_scope
grc subj_switch 0.954 87 s2:2611:v1.2 both_named

Features not listed have not been checked this way. In particular the argument roles of has_argument and the linked_to targets carry no figure yet.

Not here

  • RDF: the tables are the source of truth; an RDF export is planned later.
  • Rabbinic commentary terms: in bcv-commons/hebrew-commentary-terms; not joined in this version.
  • Occurrence groups for lemmas UBS leaves without a sense (V3c): computed but not published; the groups have no labels, a two-way split can follow form rather than sense, and the commentary yardstick for them is thin.

Sources and licence

Inputs (bcv-commons datasets, revision in brackets):

  • hebrew-word-concepts-model v0.6 (dbdb2c9f)
  • greek-nt-concepts-model v0.1 (6f73ed21)
  • word-senses v0.1 (00690ee3)
  • hebrew-greek-bridge v0.1 (1bee0403)
  • hebrew-greek-lexeme-bridge-model v0.3 (ffcb5e7c)
  • hebrew-nt-bridge v0.1 (525db08f)
  • clause-columns v1.2 (32beff0b)
  • argument-roles v0.2 (5065acab)
  • time-aspect v0.2 (5d30d2c2)
  • clause-connectives v0.1 (c26a84be)
  • the bridge corpus hebrew-greek-cf (build b15)

Underneath: MACULA Hebrew and Greek (CC BY 4.0), the UBS dictionaries (CC BY-SA 4.0), GLAUx (CC BY-SA 4.0), ACAI (CC BY-SA 4.0), the Hebrew NTs of Salkinson-Ginsburg and Delitzsch (public domain). Nothing comes from BHSA or TBTA. This dataset: CC BY-SA 4.0.

Provenance

Code 341e013: shoresh/macula/build_source_ontology.py, shoresh/macula/build_ontology_propositions.py, shoresh/macula/release_source_ontology.py (with release_entity_mentions.py for the entities).