Analysis

How far 154 transbliterations travel from Shakespeare's sonnets, and from the Spanish they were written from.

Introduction

Each poem in Transbliterations: Shakespeare began as one of Shakespeare’s 154 sonnets. The sonnet was put through Google Translate into Spanish in 2010, and the poem was written from that Spanish, read partly for what it meant and partly for what it sounded like in English. The History page tells how the method came about, and notes that others had used related methods before: homophonic translation, erasure, and poems made from machine translation’s mistakes.

This report asks what the process left behind. For each poem, how much of it

• restates what its sonnet says, in Shakespeare’s words or in modern ones;
• takes his words, sounds or images, or the Spanish’s, and makes them say something else;
• says the opposite of the sonnet;
• has no source in either?

It also asks how much of each poem came from something only the Spanish offered, and how good that 2010 translation was.

Methods

Texts. The 154 poems as they stand in the book; the sonnets they answer, as printed in 1609; and the Spanish each poem was written from, copied from the working document of the time. The book arranges the poems in six Parts, in an order that is not Shakespeare’s.

Word counts. A script compares each poem with its own sonnet and, as a baseline, with the other 153:

• lift, how much more the poem’s words overlap its own sonnet than the others;
• aligned, how many sonnet lines are matched in order by runs of poem lines, against chance;
• order, whether the matched lines keep the sonnet’s sequence;
• verbatim, the words inside runs of three or more words shared with the sonnet;
• Spanish look-alikes, poem words that resemble a word of the Spanish but not of the English.

These see shared words, not paraphrase or sound, so they are a check on the reading below rather than a replacement for it.

The line judge. An AI model (Claude) acted as judge. It read every poem line by line beside its sonnet and the Spanish, 4,005 lines in all, and answered three questions for each line, following a written rubric with seventeen worked examples:

1. How did Shakespeare’s English arrive: verbatim, through a shared word or image, through sound, or not at all?
2. Did the Spanish shape the line: faithfully, or through a sound, a cognate, a mistranslation or untranslated English? The judge had to quote the Spanish words responsible.
3. What does the line do with its source: make the same claim, a shifted one, the opposite, or something new?

The judge was not told the book’s Parts, when the poems were written or anything about the poet’s life, and it worked through the poems in shuffled batches, so no pattern across the Parts could be suggested to it. It also wrote a short reading of each poem, which the poem pages show as Claude’s reading.

From labels to shares. A script, not the judge, turns the three answers into one label per line and the labels into four shares: Restated, Transbliterated, Inverted and Original. Shares are word-weighted. When a transbliterated line keeps words from its source line, half of those words’ weight moves to Restated (kept words), on the rule that a shared word makes a line slightly more of a restatement.

A second, separate question. Three more shares record where each line’s words and images came through, whatever the line then does with them. Faithful lines took them from a part of the Spanish that translated Shakespeare correctly, Noise lines from something only the Spanish offered, and Untouched lines took nothing through the Spanish. A Faithful line can still change or reverse Shakespeare’s claim, so these shares do not line up with the four above.

The translation judge. A second judge rated each line of Google’s Spanish for adequacy and fluency, tagged its errors, and wrote a faithful translation of its own, which chrF compares with Google’s.

Checks.

• Every batch of judging labeled the same two anchor poems again.
• A fresh judge labeled ten randomly chosen poems a second time (blind re-judging).
• Sonnet I is the one poem whose drafts survive, and they trace 20 of its lines to their sources.
• The judge’s shares were compared with the word counts by Spearman ρ.

Results

The book

A quarter of the words (25%) restate Shakespeare. Just over half (56%) are transbliterated: his material, or the Spanish’s, turned into another claim. A tenth say the opposite of the sonnet (9%), and a tenth have no source (10%).

The two bars answer different questions, which is why 59% of the words can be Faithful when only 25% are Restated. Only a quarter of the Faithful words restate Shakespeare; two-thirds are transformed. In Sonnet I, “Pity the world, or else this glutton be” came through the Spanish intact, and the poem made it “Here is your glutton!”: the glutton is Shakespeare’s, and the claim is not.

Reversals run through the whole book. The judge recorded 273 of them, and only 22 of the 154 poems have none. No poem makes its sonnet’s whole argument: on the judge’s scale of argument kept, 66 poems share the sonnet’s situation but not its claim, 67 keep part of the claim, 20 keep the claim with a changed emphasis, and one argues something unrelated. The poems still walk through their sonnets. Runs of poem lines match a median of 57% of the sonnet’s lines, against 9% by chance, and in half the poems at least 86% of the matched lines keep the sonnet’s order.

A quarter of the words (25%) took something only the Spanish offered. Mistranslations taken at their word are the largest part (13%), then look-alike words (7%), then sounds and untranslated English (3% each). In Sonnet XC, Shakespeare writes “do not, when my heart hath ‘scap’d this sorrow, / Come in the rearward of a conquer’d woe.” In the Spanish, the “no” stays on the first line and the second says Ven en la trasera, come in behind. The poem says “Sneak in / Please / Behind my armored woe,” which the judge read as a mistranslation taken at its word, and as an inversion.

The translation

Google Translate in 2010 carried about half of Shakespeare’s meaning (adequacy 2.0 of 4) into Spanish that read poorly (fluency 1.5 of 4). Grammar errors appear in 78% of lines and a word in the wrong sense in 48%. How badly a sonnet was translated had little to do with how far its poem went: the correlation of a translation’s adequacy with its poem’s departure from the English is ρ = −0.21, and with the poem’s share of Spanish noise ρ = −0.08.

Across the Parts

From Sensing to Rebelling, the poems restate more of Shakespeare: 20% of a Sensing poem on average and 33% of a Rebelling poem, with intervals that do not overlap. Over the same Parts, Transbliterated falls from 62% to 50% and Spanish noise from 30% to 22%. Inverted stays near a tenth throughout. The last Part, Accepting, holds five poems, and its figures return to Sensing’s.

Across the days

The book is ordered by the day each poem was written, counted from day 0, so the same rise can be read against time. The 17 poems written before day 0, all of them in Sensing, restate 21% of their sonnet on average. The 83 written from day 0 to day 419 restate 24%, and the 54 written from day 420 on restate 31%, with the least Original, 7%. Across all 154 poems, the later a poem was written, the more of its sonnet it tends to restate (ρ = 0.35). The day axis also shows the pauses: just over 600 days separate Sensing’s first eight poems from its last nine.

Checks

• A poem’s own sonnet is its closest word match for 150 of the 154 poems. The four exceptions are candidates for a mislabeled numeral.
• The judge’s Restated share rises with lift (ρ = 0.65). They cannot agree fully, because a modernized line may share no words with its source.
• The judge’s verbatim share agreed less with the verbatim word runs than expected (ρ = 0.50, where 0.7 or more was expected).
• The share the judge traced to Spanish noise rises with Spanish look-alikes (ρ = 0.33). Look-alikes cannot see sounds or mistranslations, so this is a floor.
• chrF follows the judged adequacy (ρ = 0.55), and the English that a script counts as left untranslated agrees with the judge’s tags for it (ρ = 0.89).
• The judge agreed with Sonnet I’s drafts on all 20 traced lines.
• The two anchor poems, labeled again in eleven batches each, moved at most 7 and 13 points on any share.
• In the ten poems labeled twice, 86% of words landed in the same share, shares differed by 4.3 points on average, and argument kept agreed within one point every time.
• Of the judge’s 273 reversals, 209 have no negation or antonym pair a script could find, which is why reversals were judged rather than counted.

The poems on this site

Conclusions

• Most of each poem is neither Shakespeare restated nor free invention. Over half of the words take his words, sounds and images, or the Spanish, and turn them to another claim; a quarter restate him; a tenth reverse him; a tenth owe him nothing.
• The Spanish mattered. A quarter of the words came from what only the translation offered, most often a mistranslation taken at its word.
• A worse translation went with only slightly more departure from Shakespeare, and not with more use of the Spanish’s noise.
• From the first Part to the fifth, the poems come slightly closer to what Shakespeare’s sonnets say.

These figures have limits. The line labels are an AI’s reading by a written rubric, checked as above but not confirmed line by line. The links to the Spanish are readings of the final state of the poem rather than records of the process by which any line was written by the author; only the drafts of Sonnet I show that, and they agree.

Glossary

Two questions

From the English

What each line does with Shakespeare's claim: keeps it (Restated), turns it into another claim (Transbliterated), reverses it (Inverted), or has no source (Original).

From the Spanish

Where each line's words and images came through: a part of the Spanish that translated Shakespeare correctly (Faithful), something only the Spanish offered (Noise), or not the Spanish at all (Untouched). A separate question from the English one, so the two don't line up: most Faithful lines still change Shakespeare's claim.

The four shares

Restated

Shakespeare's claim kept, in his words or in modern ones: every verbatim or modernized line, plus half the weight of the words a transbliterated line keeps from its source line.

Transbliterated

Shakespeare's words, sounds or images, or the Spanish, turned into a different claim: the transformed, sound-echo and spanish-echo lines.

Inverted

A line that says the opposite of its source line: a negation, praise turned to contempt, lasting turned to decaying. A single reversed word counts.

Original

A line with no source in the sonnet or in its Spanish.

word-weighted

Each line counts by its number of words, so a one-word line weighs less than a long one.

kept words

When a transbliterated line keeps words from its source line, half of those words' weight moves to Restated.

From the Spanish

Faithful

The line's words or images came through a part of the Spanish that translated Shakespeare correctly, so nothing in them is unique to the Spanish. This says where the material came from, not what the line does with it: a Faithful line can still change or reverse Shakespeare's claim.

Noise

The line's words or images came from something only the Spanish offered: a sound, a look-alike word, a mistranslation, or English the translator left untranslated. Like Faithful, it says where the material came from, not whether the claim was kept.

Untouched

Nothing in the line came through the Spanish: mostly lines with no source at all, and lines that took an image or a sound straight from Shakespeare's English.

sound

A Spanish word heard as a different English word, as llamativas (gaudy) heard as "call".

cognate

A Spanish word read as its English look-alike, false friends included, as aumentan read as "augment".

mistranslation

The line follows a meaning Google Translate got wrong, as Shakespeare's "waste" coming back as residuos, residue.

untranslated

The line uses English that the translator left standing in the Spanish.

Line labels

verbatim

Three or more consecutive words of the source line survive, or half its content words, and the claim is the same.

modernized

The source line's claim in new words: a line a teacher would accept as a loose modern translation.

transformed

Shakespeare's words or images kept, but used to make a different claim.

sound-echo

A word that sounds like one of Shakespeare's but means something else (spite → despite), used for a different claim.

spanish-echo

A line shaped by something only the Spanish offered (see Noise), used for a different claim.

S, T and P

S5 is line 5 of Shakespeare's sonnet, T5 the same line in Google's Spanish, and P5 line 5 of the poem. "No source" means neither.

Word counts

lift

The share of the poem's content words found in its own sonnet, minus the average share found in the other 153 sonnets. Zero is no more overlap than chance.

⚑

The poem's words match a different sonnet better than their own, which can mean a mislabeled numeral.

aligned

The share of the sonnet's lines matched, in order, by a run of one to four poem lines that beats 95% of matches against other sonnets.

chance

The same measure taken against other sonnets, which the poem was not written from.

order

Of the poem lines that match a sonnet line, the share that follow the sonnet's order. 1.00 means the poem never jumps back.

verbatim

The share of the poem's words inside runs of three or more words shared with the sonnet.

Spanish look-alikes

The share of the poem's content words that resemble a word in the sonnet's Spanish but appear nowhere in its English.

left in English

The share of the Spanish translation's words that Google left in English.

The translation

adequacy

How much of Shakespeare's line the Spanish carries, from 0 to 4: 4 is all of it, 2 about half, 0 nothing or the opposite.

fluency

How well the Spanish reads as Spanish, ignoring the English, from 0 to 4: 4 is natural, 2 awkward, 0 word salad.

chrF

A character-level match score from 0 to 100, comparing Google's Spanish with a faithful translation written by the translation judge.

grammar

Agreement, conjugation or word order broken.

wrong sense

A word translated in a sense that doesn't fit, as "tender" becoming licitación, a bid.

untranslated

English left standing in the Spanish.

archaic misread

An Early Modern word or sense taken as a modern one.

omission

Meaning dropped.

addition

Meaning added.

role swap

Who does what reversed, as "we desire increase" becoming creatures that increase desire.

The judge's readings

judge

An AI model (Claude) that read each poem line by line against its sonnet and the Spanish, and labeled every line by a written rubric. It never saw the word counts, the Parts or the poet's story.

Claude's reading

Sentences the judge wrote: its summary of the poem, of the sonnet's argument, and its notes on single lines. They are an AI's reading, not the poet's account.

The sonnet

The judge's one-sentence summary of what the sonnet argues.

argument kept

How much of the sonnet's argument the poem keeps, from 0 to 4: 0 unrelated, 1 the same situation but not the claim, 2 part of the claim, 3 the claim with changed emphasis, 4 the sonnet's whole argument.

feeling

The main target of feeling in each text and its valence, from −2 (hostile or despairing) to +2 (adoring or hopeful), then how it moved: none, intensified, softened, reversed (crossed zero), redirected (a new target) or dissolved (no clear stance).

spoken to

Whom each text addresses.

confidence

The judge's own confidence in its line labels for the poem, low, medium or high.

quadrant

Where the poem falls on two splits: lift above or below the median of all 154 poems (form), and Restated above or below half (sense). Close paraphrase, new-words paraphrase, reworked borrowing, or free-standing.

axis

The kind of reversal: love-hate, praise-blame, permanence-decay, hope-despair, affirmation-negation (yes and no), or other.

darker / brighter

Whether the poem's side of a reversal is darker or brighter than the sonnet's.

Δ valence

The poem's valence minus the sonnet's, from −4 to +4. Negative is darker. For a Part, the average over its poems.

Statistics and checks

Part figures

The average of each poem's shares within a Part, so every poem counts equally. The book-wide figures weight by words instead.

90% interval

The range holding 90% of 2,000 bootstrap resamples of a Part's poems: how much the figure depends on which poems the Part happens to contain.

running average

Each point averages the shares of nine neighboring poems in book order.

day

Days are counted as the book counts them, before and after the day 0 that the History page describes. A poem sits on the first day the book lists for it, when it was written, and is grouped with the other poems written in the same 30 days. Later days listed for a poem are revisions.

Spearman ρ

A rank correlation from −1 to +1: +1 when two measures put the poems in the same order, 0 when their orders are unrelated.

anchor poems

Two poems that every batch of judging labeled again, to check that the batches agreed with each other.

blind re-judging

Ten poems labeled a second time by a fresh judge, to see how often the same words land in the same share.