alignments-by-id — translations word-aligned to MACULA Greek word ids

New Testament translations word-aligned to the Greek words they translate. Each Greek word is identified by its id in the MACULA Greek (Nestle 1904) trees (n40001001001 = Matthew 1:1, word 1), so a link can be joined to the trees' syntax directly. A translation made from the Textus Receptus (the KJV) is aligned to the TR instead (tr/), and each link also carries the Nestle 1904 ids of the same words where the two texts agree. Where the translation writes a function word inside another word (Hebrew ו ה ב ל מ כ ש, pronoun suffixes), the link points at that part of the word.

These are machine alignments. Each link carries a class whose hand-checked precision is in the manifest; use the classes to choose your own precision.

Editions

folder	edition	Greek words linked	not rendered (none)
n1904/h/heb/heb_scg	Salkinson–Ginsburg Hebrew NT (public domain; bible.helloao.org, eBible hebsg), 27 books	118,166 of 137,779	19,580
tr/e/eng/eng_kjv	King James Version, 1769 text (public domain outside the UK; bible.helloao.org eng_kjv = eBible eng-kjv2006), 27 books	129,513 of 140,856 TR words	11,343

The KJV's Old Testament is in full-alignments-llm (e/eng/eng_kjv+llm+jev-v4, keyed like compact-alignments to MACULA Hebrew).

The Shorter Ending of Mark (MACULA MRK 16:99, 33 words) is not in Salkinson–Ginsburg; it is listed in the manifest (spine_verses_not_in_edition) and appears in no file.

The Greek side

n1904/ holds alignments to the Nestle 1904 text of MACULA Greek (Clear Bible, CC BY 4.0). An id is n + book (40–66) + chapter (3 digits) + verse (3) + word (3), exactly the xml:id of the word in the trees; all 137,779 ids were checked against the trees. n1904/manifest.json → spine pins the word list the alignment was made on (spine_content_sha256).

tr/ holds alignments to the Textus Receptus spine of alignments-tr: an id is that spine's word key (grc-tr:MAT:1). A link has a 4th element: the MACULA Nestle 1904 ids of the same words, where TR and Nestle have the same word in the same place of the same verse (longest common subsequence of the accent-free forms per verse; 119,903 of the 129,513 linked TR words have one). A TR-only reading, or a verse numbered differently in Nestle (2 Corinthians 13:14 is Nestle's 13:13), gets []. tr/manifest.json pins both word lists (spine, nestle_spine).

Versions

Each release is a git tag of this repo (v1.0, …); pin a tag (revision="v1.0") rather than main. A tag is never moved: a rebuild becomes a new tag.

File format

n1904/<iso[0]>/<iso>/<edition>/<BOOK>_<hash>.json
{"C:V": {"links": [[[positions], [ids], class], ...], "none": [ids], "added": [positions]}, ...}

• C:V is the verse in the edition's own numbering. When the edition joins verses, a verse can hold ids of the following verses.
• A position is a token of that verse, by the tokenizer of compact-alignments (tokenize.js, tokenizer v2, from bcv-commons/bibles verse-text): the integer 6 is the whole token, the string "6.0" is part 0 of token 6. Part numbers are those of the word analysis named in the manifest (parts: bcv-commons/hebrew-nt-analysis v0.1, words split only when the analysis is sure, p_split ≥ 0.9). Part letters can be read from that dataset.
• One link may hold several ids when adjacent Greek words with the same Strong's number were aligned as one unit.
• none: Greek words of the verse that the translation does not render. added: tokens that render no Greek word.
• supplied (KJV): tokens the edition prints in italics, words the translators supplied (from the \add markup of eBible's USFM). A supplied word may still be linked to the Greek word it completes.
• A verse's leading verse number (written in Hebrew letters, א, יז …) is a token too; it is never linked and never added.
• <hash> is the book's text fingerprint, the same name the compact-alignments file of that text has, so a file only matches the exact text it was built on.

Classes

class	what it means	hand check (heb_scg)
kept	the aligner's link, which the reviewer (Jev) kept	25/25
inside	Jev: a function word rendered inside a word another Greek word owns; the part is the aligner's, or the one part whose class fits	21/25
part	Jev said "not rendered", but the aligner linked the word to a prefix or suffix whose class fits it (conjunction → ו, preposition → ב ל מ כ, article → ה, pronoun → suffix)	21/25 (47/61 in a separate comparison)
edited	Jev changed the link, one target word	25/39
edited-wide	Jev changed the link, two or more target words; often a neighbouring word of another Greek word swept in	17/36

none was right (the word really has no rendering) in 19/25 checked cases.

KJV (tr/), checked against the KJV's own Strong's tagging (eBible eng-kjv2006), on Greek words whose Strong's number occurs once in the verse and is tagged: kept lands on the tagged words 98.0% (exactly them 91.3%, n 59,921), edited 88.8% (52.6%), edited-wide 97.3% (16.4%: the tagging marks the head word, the link also holds its articles, auxiliaries and pronouns), inside 93.7% (90.8%). Of the none words, the tagging gives 47.8% a word (n 3,771): treat none on the KJV as a weak claim. The tagging was also an INPUT: a word whose proposed link matched its tag was not sent to the reviewer, so kept's figure is partly by construction.

Greek words on 2+ target words: 7,428, of which 5,316 are edited-wide (see its precision) and 2,112 kept.

How the links were made

1. The standard lexeme-aligner chain (eflomal, gloss, span extension, gap fill) on a parts view of the edition: each word the analysis splits is given to the aligner as its parts, then the links are folded back to tokens and parts.
2. Jev (TypeSafe), a reviewer model, checked the proposed links of every verse (design v4, threshold 0): keep, change, rendered inside another word, or not rendered. For the KJV, the edition's Strong's tagging cleared a word's doubt when the proposal matched it, and frequent function-word links Jev kept at least 97% of the time on the first books were not asked again (--skip-pairs). The conventions text the reviewer was given for each book is in that edition's _conventions.json and in alignment-conventions.
3. The part rule, measured on 130 hand-graded cases where Jev said "not rendered" for a prefix the aligner had linked.

Code: lexeme_aligner.alignments_by_id (github.com/bcv-commons/lexeme-aligner).

License

The alignments are CC0 1.0. They hold no text; each edition keeps its own licence (Salkinson–Ginsburg: public domain). MACULA Greek (CC BY 4.0) supplies the ids; hebrew-nt-analysis (CC BY 4.0) the part numbers. No other value from either is copied into the files. The KJV: TR spine from alignments-tr (CC0); the text is public domain outside the United Kingdom (Crown letters patent restrict printing there).