Bible As Written

Our process

How we turn named biblical witnesses into English you can read aloud, and how we intend to mark the same idea across Hebrew, Greek, Aramaic, Syriac, Ge’ez, and Latin without pretending verse numbers or English word order are the meaning.

Layers we actually store

We do not store “one verse row per language” as the real text. A verse number is a citation window (usually Masoretic / common English breaks). The text itself is an ordered list of words (tokens) for each named witness.

Close English (per witness)

Close is three stages: gloss each form, assemble English, repair only when the assembly is nonsense. We did not invent “gloss then light edit.” Field linguistics and Bible-translation manuals already separate a word-by-word gloss from a readable line.

What we do invent at this layer is the product cut: Close must stay near the gloss, Clear may read easier, and neither may sand off a hard image the source actually has.

Matching ideas across languages

Hebrew often puts the verb first. English Close puts God before “created.” Greek, Ge’ez, and Latin each have their own order. So we must not say “the sister is the next English word in the line.” That is how a leftover YHWH gets paired with a leftover the.

Pass 1 — same idea. If two content-words share a gloss (or a known same idea: heavens / heaven, God / Egziabher), lock them even if they sit in different places. Order is only for “first earth” vs “second earth.”

Pass 2 — conflict or gap. What is left may still be one slot with two wordings (created / made). We keep that link and mark it as a conflict. A witness with nothing in that slot is absent (Ge’ez and “host”). We do not zip leftover bags in stream order.

Verses stay for navigation: “open Genesis 2:1.” They are not the atom we align.

What we take from prior work

Other people have aligned meaning across languages for decades. We are not pretending this is a blank page. These are the pieces we actually use as warrants:

A lexicon first, not word position

IBM Model 1 (Brown, Della Pietra, Della Pietra, Mercer 1993) aligns words by which word generated which, with every position equally likely. Later models add “stay near the previous link.” That helps French–English. It hurts Hebrew vs English Close.

Brown et al., “The Mathematics of Statistical Machine Translation,” Computational Linguistics 19(2), 1993. ACL Anthology

Och & Ney, “A Systematic Comparison of Various Statistical Alignment Models,” Computational Linguistics 29(1), 2003. They also show how to score a method against a hand gold alignment. ACL Anthology

Dyer, Chahuneau & Smith, “A Simple, Fast, and Effective Reparameterization of IBM Model 2,” NAACL 2013 (FastAlign). Turning up the “keep the same order” knob prefers monotone lines. ACL Anthology

Windows vs words; missing is allowed

Gale & Church (1993) align sentences by length and treat 1–0 and 0–1 (no counterpart) as normal. A Bible verse is that kind of window, not a proof that every word has a sister. IBM “NULL” / fertility is the same idea at word scale: one language can omit or double.

Gale & Church, “A Program for Aligning Sentences in Bilingual Corpora,” Computational Linguistics 19(1), 1993. ACL Anthology

Kay & Röscheisen, “Text-Translation Alignment,” Computational Linguistics 19(1), 1993 — partial word anchors first, then the rest. ACL Anthology

Hebrew and Greek already did plus / minus / different rendering

The CATSS project (Emanuel Tov and colleagues) built a parallel MT–LXX database: formal equivalents, pluses, minuses, and cases where the Greek is not a straight match to the Masoretic line. That inventory — same slot, different wording, extra, missing — is what we want for six witnesses, not two.

Tov, “The Use of a Computerized Data Base for Septuagint Research,” BIOSCS 17 (1984); CATSS vol. 2 (1986). Intro: CCAT parallel README

Gloss hangs on the source word, not on the free English

The Leipzig Glossing Rules align a gloss with the example in the source language. The idiomatic translation is a separate line. Close is our strict readable line. Matching two Close lines by English order is matching two free translations, which those rules never asked for.

Comrie, Haspelmath & Bickel, Leipzig Glossing Rules (2008 / 2015). MPI EVA. They cite Lehmann, “Directions for interlinear morphemic translations,” Folia Linguistica 16 (1982).

Same meaning, different grammar

Bible-translation manuals already allow “modified literal” English and warn against both interlinear mush and undue freedom (Beekman & Callow; Larson; Nida’s formal vs dynamic scale). Catford’s translation shifts are the name for “same situation, different grammar.” That is a conflict in one slot, not two unrelated leftovers.

What we are inventing here

These parts are ours as a product, not a claim that academia never aligned bitext:

• Two English layers (Close vs Clear) with an anti-smoothing rule on hard images.
• Style locks such as YHWH and Egziabher that must survive Close assembly.
• Named-witness stack in one reader (including DSS scraps that do not pretend to be a full scroll).
• N-way slots with an explicit absent member (six Closes, not a pair), stored on tokens, shown in the sister list.
• Pass 2 rule we are adopting: leftover conflict only by a listed pair or a shared already-locked neighbor — never by “next leftover in the stream.”

We will check the slot matcher against a small hand list (Genesis 1:1 created/made; Genesis 2:1 Ge’ez missing host; Genesis 2:7 YHWH is not “the”), the same way Och & Ney insist on a manual gold alignment instead of trusting the algorithm’s own story.

Not only the Bible

The same three outcomes — same idea, different wording, missing clause — show up wherever a sacred or classical text moved across languages. Our first reading list was Western bitext papers and the Hebrew–Greek CATSS files. That was too narrow. Hindu scripture in Indian languages, Chinese Buddhist translation, and Japanese reading of Literary Chinese all hit this problem. They do not all use English SVO, and a lot of the useful writing is not in English.

India. Sanskrit verse is often not the order you understand it in. Commentators write an anvaya (prose construction) with the same words in a different sequence. Traditional tests for “these two words belong together” are expectancy, congruity, and proximity. A Mīmāṃsā line that computational Sanskritists still quote: those three are the grounds of a relation; mere next-word sequence is not. That is our Pass 1 / Pass 2 split, named in another school. The Mahābhārata critical edition puts extra and missing lines in footnotes and appendices — plus and minus, not “the next surviving line is the sister.” Tamil, Awadhi, Odia, and other Rāmāyaṇas retell the story (Ramanujan’s word is tellings, not copies of one original); computer projects that line them up usually line up chapters (sarga), which is our verse window, not a word slot. An Indian machine-translation project (Anusaaraka) already refused a guessed fluent line: it prints a morpheme-by-morpheme image of the source, which is the same job as Close.

China. In the late 300s Dao’an listed five ways a translation loses the original. The first is blunt: the foreign wording is all inverted to follow Chinese. That is what Close English does to Hebrew. He counted it as a loss of the source line, not as proof that leftover words match. Xuanzang’s five cases of “do not translate the sense, only the sound” (secret speech, many meanings, no such thing here, keep the old sound, keep reverence) is fail-closed naming — the same instinct as locking YHWH and Egziabher. Chinese 梵汉对勘 work (Karashima and others) collates a Han translation against Sanskrit by attested equivalents, not by sitting in the same stretch of the line.

Japan. Literary Chinese is verb-object; Japanese is object-verb. Readers mark the Chinese characters with return-marks (kaeriten) so they can read in Japanese order without replacing the source line. Our token-to-Close map is that kind of mark. The bug in the live hover is lining up two of those readings by English order.

Tibet. Chinese and Tibetan Buddhist translations of “the same sūtra” were often not made from the same Indian parent. Felbur, Meelen, and Vierthaler (2022) still publish a hand alignment score sheet before they run an engine. We will do the same for Genesis 1–2.

A warning from the geyi debate: pairing an Indian idea with a handy Chinese category because it “fits” is how you get a false sister. Mair (2012) argues the popular “Daoist matching” story overstates the history; the failure mode is still real.

Honesty: we can open Chinese and Sanskrit pages, but this is not a native-language review of Hindi, Tamil, or CNKI paywalled journals. The point is that the problem is not a Bible-only, English-only problem, and some of the oldest names for “do not zip the next word” are not in English.

Further reading

Full citations and how each paper maps to a BAW step live in the repo: resources/research/CROSS_LANGUAGE_ALIGNMENT_LITERATURE.md. The slot-matcher rewrite (specified, not shipped) is resources/research/SLOT_MATCHER_REIMPLEMENTATION.md (Close-editing manuals: GLOSS_TO_ENGLISH_EDITING_LITERATURE.md).