📜 The Tanakh in 32 English translations

Every translation is read by Mova into a world of its own — who does what to whom, when, in what order, negated or not — and the worlds are linked through one meta-world that belongs to no translation. Only the verses that all 32 translations have are compared, in blocks of whole chapters up to 2 500 words: the text is truly parallel, so no difference of canon or numbering enters the numbers. A difference is a measured fact about two texts, not a judgement on either.

What the numbers show

• Editions of one text come out as one. 4 groups of translations are at logical distance 0 from each other: the same events, roles, order, negation and numbers. What is left between them is wording and spelling, which the bits measure. They are folded into one row below:
  • eng-web-c, eng-web, eng-webbe, engwebpb, engwmbb, engwebp, engwebu, engwmb — bits between them 1606–2370 per 1 000 words
  • eng-lxx2012, eng-uk-lxx2012 — bits between them 605–609 per 1 000 words
  • eng-kjv, eng-kjv2006 — bits between them 1086–1100 per 1 000 words
  • engbsb, engmsb — bits between them 1267–1269 per 1 000 words
• JPS 1917 stands by the ASV, not by the other Jewish translation. Its logical distance to the ASV is 10, to Leeser's 41. JPS 1917 was made on the Revised Version and the ASV; Mova did not know it and measured it.
• The Septuagint is an island. Brenton and the LXX2012 are 7 apart; the nearest translation of the Hebrew is 60 away. They translate another text, the Greek, and the difference is in the events themselves, not in the style.
• Revisions stay by their source. The Literal Standard Version, a revision of Young's Literal Translation, is 23 from it.
• Douay-Rheims stands apart: its nearest neighbour is 51 away. It is translated from the Latin Vulgate, not from the Hebrew.
• The paraphrases are farthest from everyone: eng-t4t (120 on average) and engfbv (108). A translation that restates the meaning in other words changes the events and roles a reader finds, as it should.

The tree of the translations

The translations joined by average linkage (UPGMA) on the logical distance, the mean of its two directions: the nearer two texts, the further left they join. Editions of one text stand as one leaf.

One verse, two texts: Joshua 1:9

Where translations part in an event, not in a word. The Hebrew asks a question, the Greek states:

KJV (from the Hebrew): Have not I commanded thee? Be strong and of a good courage …

Brenton (from the Greek): I have commanded thee; be strong and courageous …

Douay-Rheims (from the Latin): Behold I command thee, take courage, and be strong.

Mova's comparison says: «command» negated on one side only. That is true of the words and not of the meaning: «Have not I commanded thee?» is a question that asserts. The frames read the negation and not yet the speech act; when they do, this verse will move from contradiction to agreement. The same first block also shows plain errors of the reader (in Joshua 1:10 the parser took «and» for the agent of «command»), which are counted with the rest.

Nearest neighbours

The distance in its two parts

For each translation, the mean over all the others (both directions, per 1 000 words) of its contradictions — the same event negated on one side only, another named agent, another number — and of half its gaps — events one side has and the other does not. A paraphrase and an old translation can have the same distance from very different mixtures. The reader's own errors are in both parts: a parser that reads one style worse than another adds to its distance. The last two columns say how well each text is read: on the same sample of verses and the same words, the agreement of Mova's parse with Stanza, and of Stanza with UDPipe (labelled attachment, punctuation left out). The modern paraphrases are read best, so their distance is the text's and not the reader's; the oldest translations are read worst, and part of their distance is the reader's.

Against the Hebrew

The Leningrad Codex, read into the same meta-world: bits per 1 000 words of the translation to write it out of the Hebrew's world, and back. The Hebrew's reader is learned on the training books against all the translations alike, so that none of them is its model; the control writes each translation out of the Hebrew of another verse of the same block. Only the verses whose Hebrew and English numbers agree are counted. The closer a translation stays to the Masoretic text, the fewer bits; the Septuagint's translations stand further off, because they translate the Greek.

Residual rules

For every ordered pair, the rules that bring one translation's world nearest the other's: a concept for a concept, a part of speech for another, a word dropped. Each is learned on one half of the development blocks and measured on the other, then joined to the rules before it; the list below sums them over all the pairs. The last column measures the same rule with the verses paired by the wrong number: a rule that saves as much there is a habit of the target's style (punctuation, spelling), one that saves much less is a correspondence between the two texts (the divine name, thou and you).

Inside each family

The logical distance between the translations of one family, row to column; the Protestant ones first all together, then by lineage. Editions at distance 0 are folded into their first.

Jewish (J)

Catholic (C)

Orthodox (O)

Protestant, all lineages (P)

Protestant: the Tyndale and King James line (PK)

Protestant: the revisions (RV, ASV, WEB) (PR)

Protestant: literal (PL)

Protestant: meaning-based (PF)

Against the Jewish translations

Each translation's logical distance (the mean of the two directions) to the two Jewish translations, JPS 1917 and Leeser's, nearest first.

Bits to write one out of the other

A row is the source, a column the target. A cell says how many bits it takes to write the target's verses out of the source's meta-world, per 1 000 words of the target: the fewer, the closer. The measure has a direction: the bits from A to B are what B has that A's world does not, so a fuller text written from a terser one costs more than the other way round. The shade is the cell's rank among all cells of its measure, light (near) to dark (far), one scale for every table of the page; the letter after a name is its family (see «The translations» below). Editions at logical distance 0 are folded into their first.

Logical distance

Per 1 000 words: the contradictions (a negation, a modal, a named agent or patient, a number that differ in matched events) plus half the events one side has and the other not.

Does the distance see a change of meaning?

A distance of 0 between two editions is only worth something if the distance moves when the meaning does. The control: into every target verse of every pair one change of meaning is put, on an event the source verse matches — a negation flipped, the agent and the patient swapped, a number changed, two events reordered — and the comparison is made again. «As contradiction» is what the logical distance counts; «by any part» adds the gaps, the shared roles and the order, which the comparison reports beside it. What it does not catch is the next work: a swap of two roles whose words differ already between the two translations is not seen as a reversal, and the order of events is reported but not counted in the distance, since translations reorder clauses freely.

The translations

The family is the tradition each edition serves and, for the Protestant ones, its lineage: J Jewish, M Messianic, C Catholic, O Orthodox, PK Protestant: the Tyndale and King James line, PR Protestant: the revisions (RV, ASV, WEB), PL Protestant: literal, PS Protestant: modern standard, PF Protestant: meaning-based. On the shared verses the tradition says nothing of the text, which «translated from» does. eng-web-c is the World English Bible in the Catholic canon: on these verses it is the WEB.

The texts are from eBible.org, each under its own licence (most public domain; the share-alike ones are used for measurement only). Measured on the development books — Joshua, Nehemiah, Ezekiel, Zephaniah; the test books (Numbers, 1 Samuel, 2 Chronicles, Psalms, Jeremiah, Joel, Nahum, Malachi) are not read. The same run, two times, gives the same numbers.

🐙 The snake judges no one; it only trolls (in older days this kind of joke was called Satan: the adversary).

This page in Markdown → https://md.mova.elamur.ai/free/en/holy-wars/tanakh/