larsgson's picture
entity-mentions-model v0.1 (code 014b4da)
fd066be verified
|
Raw History Blame Contribute Delete
3.8 kB
metadata
license: cc-by-sa-4.0
language:
  - hbo
  - arc
  - grc
pretty_name: Entity mentions (model)
tags:
  - bible
  - hebrew
  - greek
  - coreference
  - model-generated

entity-mentions-model v0.1

Made by a language model (claude-opus-5-5, effort medium). No person has checked it.

For pronouns that bcv-commons/entity-mentions v0.1 leaves without an entity (neither ACAI nor MACULA's reference chain reaches them; bcv-commons/reference-status v0.2 lists them as unresolved), the entity the model reads them as referring to, or why it gives none. Use it beside entity-mentions, not in place of it.

Scope. Personal, demonstrative, reflexive, reciprocal and relative pronouns and pronoun suffixes. Interrogatives (מָה, מִי, τίς), indefinites (τὶς) and words of quantity or quality (ὅσος, τοιοῦτος) are left out: they refer to no entity (Hebrew 1,228, Greek 1,159 words).

No English. Entities are entity-mentions' neutral codes; reasons are codes.

How it was made

436 calls, one per chapter (up to 25 words each). The model got, in Hebrew or Greek only: each word's verse and the three verses before it, the words marked, every word entity-mentions links to an entity tagged with a candidate number, and the candidates (the entities entity-mentions has in this chapter and the one before: type, gender, lemmas). It picked one candidate per word or none with a reason. An answer naming no listed candidate is dropped. Run cost USD 13.46 (reported by Claude Code; the user's subscription).

Table (data/answers.parquet, 1,612 rows)

column
key MACULA word key (Tanakh 12 digits, Greek NT Nestle 1904 11 digits)
lang hbo or grc
entity the entity code (entity-mentions), or empty
none_reason where entity is empty: proposition (it points at a statement, event or matter), not_listed (an entity that is not among the candidates, often one entity-mentions does not have, such as "someone" or a group the text gathers), unclear
candidates how many candidates the model had
Hebrew and Aramaic Greek NT
words 434 1,178
an entity 79 159
none: proposition 52 471
none: not listed 301 544
none: unclear 2 4

How far it agrees with blind judges

A sealed sample (seed 1119, 60 words per language, drawn before scoring; rules registered before judging) was judged blind by Claude subagents, two per word, each from the word's verse and the three before it with the same candidates. Judges are the same model family, so this measures agreement, not truth. Scored on the choice (candidate or none), once.

Hebrew Greek together
words 60 60 120
model = judge A 0.917 1.000 0.958
model = judge B 0.917 1.000 0.958
judges agree 0.950 1.000 0.975
model = judges, where they agree (n) 0.947 (57) 1.000 (60) 0.974 (117)

Most answers on every side are none, which makes these figures easy to reach. Where the model or a judge chose an entity (17 words; a count not registered beforehand), all three named the same entity for 11; the others were Hebrew words, most of them in the Psalms and Isaiah.

Sources and licence

bcv-commons/entity-mentions v0.1 (CC BY-SA 4.0; ACAI, © Mission Mutual / BiblioNexus, CC BY-SA 4.0) for the entities and candidates; bcv-commons/reference-status v0.2 (7612f663) for which words are unresolved; MACULA Hebrew and Greek (CC BY 4.0) for the text and keys. Nothing comes from BHSA or TBTA. This dataset: CC BY-SA 4.0.

Provenance

Code 014b4da: shoresh/macula/pronoun_model.py (prepare, run, collect, sealed sample), shoresh/macula/release_entity_mentions_model.py. Raw answers sha256 998b8a50a10a (in build.json in full).