A second study

Continuing her voice

Working draft · October 2026 · revised 18 October 2026 after an independent review · one model

Branches of one saved machine reading studied different amounts of Sophie Duvernoy's earlier English and were asked to continue it. In this workflow they adopted some of her earlier renderings and editorial conventions, while several structural choices persisted from the shared reading's plan. Repeated runs also differed on the instrument's categories. The study measures these similarities and differences in one chapter; it does not evaluate literary quality or establish a general limit on imitation.

The question

What changes when a machine translator studies a human translator's earlier chapters and is asked to continue in her voice?

The first study on this site compared six machine translations of Chapter 25 ("Spring") of Gabriele Tergit's Effingers with Sophie Duvernoy's published translation (New York Review Books, 2025). The six machine versions agreed with one another far more often than any of them agreed with hers; the Findings set out the measurements.

Here, branches of a single machine reading of the novel read one, three, ten, twenty or twenty-four chapters of her English, took notes on them, and received an explicit instruction to continue in that voice. The comparison measures this combined workflow (her chapters, the study of them and the instruction together), not the effect of reading alone. The question is which of her renderings and conventions then appear in the machine's Chapter 25, which do not, and whether the machine versions still resemble one another more than they resemble hers.

Design

The model is Claude Opus 5.5 at maximum reasoning effort, with file tools only (read, write, edit, search): no internet, no research, no persona. All runs were made between 13 and 15 October 2026.

The key line of the reading brief:

Translate Chapter 25 into English. You will receive precise translation instructions after the reading.

The two sentences added for doses 1 to 24, exactly as given:

The reference translation is the established English of this book. Translate Chapter 25 so that it continues seamlessly in that voice — as if you were the same translator carrying the book forward into the next chapter.

The number of her chapters a branch read is its dose. Dose 0 received only the translation instructions. The chapters are nested: each larger dose contains the smaller ones.

What every branch inherits. The shared reading was long enough for its conversation to be compacted once, so the branches start from a compacted conversation and the notes it saved. Those notes already contain a developed plan for Chapter 25: keep the verbless fragments verbless, keep the present-tense passage in the present, a fixed English wording for each return of the refrain, and a checklist to run after drafting. The notes call the plan provisional, and no branch was told to keep it or to change it. The experiment therefore asks whether her chapters and the instruction move an already developed reading and plan; it does not test imitation from an unsettled reading.

Repeats. Dose 0 and dose 20 were each run three times from the identical saved reading. At dose 20 each repeat includes studying her chapters, taking notes and revising, so the variation between its runs is variation of that whole workflow. Doses 1, 3, 10 and 24 were run once; dose 24 was added on 15 October, branched from the same saved reading.

A reference run. The first study's control brief and step messages were also run verbatim on the same model, with a reading of its own. It appears below as C0.

How it was measured

Three measurements, each against her Chapter 25 as printed in the book.

1. The first study's 45 points. The 45 places in the chapter where the German leaves a real choice, as coded in the first study, were coded for every October run, in two batches. The coding was blind: the coder saw the texts under neutral labels. In each batch two of the first study's versions were coded again as unannounced calibration anchors, and all four matched their earlier coding at all 45 points.

2. A new instrument, built blind. The 45 points were chosen to compare differently briefed machines, not to trace her earlier chapters. For this study a separate AI instance built an instrument from the German and her translation alone; it never saw a machine translation. It lists 240 points in Chapter 25 where renderings can be compared: realia and period terms (81), places and street names (27), honorifics and forms of address (24), idioms and set phrases (24), names (19), titles and professions (19), dialect and register (14), recurring phrases (13), songs, verse and quotations (10), foreign words (9). For each point it records where her earlier chapters show how she handled the same item, or a comparable one, and classifies whether her Chapter 25 repeats all of those examples, some of them, or none. It also records the chapter's structure: fragmentary sentences, tense, the refrain, the paragraphing. The instrument was finished before the repeat runs were made; when dose 24 was added, a second instance, given only the German and her translation of chapters 1–⁠4 and 25 and the instrument's list of points, added examples from her chapters 1–⁠4. It is not published, because it quotes her translation at length.

After the independent review (see the note at the end of the page), every earlier example was attached to the option it instantiates, across her chapters 1–⁠24, and 111 points were re-examined for the kind of evidence each example is: the same item, a comparable item in a closely comparable context, or a weaker analogy. The instrument's examples and its classifications of her practice are AI judgments that no person has validated.

A translation follows an earlier example at a point when the option it uses matches at least one documented earlier rendering in her chapters 1–⁠24. Whether her own Chapter 25 repeats all, some or none of those examples is a separate question, answered by the instrument's classification.

Scoring. A script matched each text to the instrument's options, testing them in the order the instrument gives, wherever the wording decided it (4,132 decisions across eighteen texts). A blind coder decided the rest (188) and coded the fragments and the tense, in two batches of anonymized texts: first the nine dose runs made by 14 October and C0, then dose 24 and the first study's six. Both batches included her own Chapter 25, unannounced, and the second repeated dose 20 as a check. Four agreement checks between procedures (not estimates of accuracy, since both procedures can share a mistaken category):

• her own chapter takes its designated option at 240 of 240 points in both batches;
• coder and script agreed at 84 of 88 and 71 of 72 points in the hidden audit samples;
• the coder's reading of her chapter's fragments matched the instrument builder's at 60 of 62 in each batch;
• on the two texts coded in both batches, the coder made the same call at 75 of 76 items (her chapter) and 75 of 78 (dose 20); the 50 fragments common to both batches' samples were coded identically.

Two versions share a category at a point when the instrument codes both with the same option. A shared other counts as a shared category although it can cover different wordings; the shared other codes are reported beside the counts. Structure is compared with her chapter as the same coder read it.

3. A style fingerprint. Burrows's Delta on the 150 most frequent words, a model-free measure from authorship attribution (lower is closer), computed across all the texts in the set: a corpus-dependent distance over frequent words, not a test of whether a reader hears her voice. Surface markers are counted per 1,000 words.

Results

Dose 0 and dose 20 are shown as the mean of their three runs, rounded; the single runs are in Table 6.

Following her earlier examples

Rows: a descriptive grouping by the chapter of the point's most recent earlier example; each point is counted once. A point counts when the option used has at least one documented earlier rendering, which may sit in an earlier chapter than the row's. Shaded: the row's chapters are among the chapters the dose read. Her own Chapter 25 uses such an option at 115 of these 155 points; C0 at 94.

Using an option with a documented earlier rendering rises from 90 of the 155 points at dose 0 to 118 at dose 20 (single runs: 89–⁠92 and 117–⁠119). The one run given all twenty-four earlier chapters reaches 115. Dose 0 already reaches many of her documented renderings without reading any of her English: they are also the renderings a translator without her chapters chooses.

A gain: the run uses an option with a documented earlier rendering at a point where none of the three dose-0 runs does. A loss: the reverse, where all three dose-0 runs do. Attested inside its reading: the option's documented rendering occurs in one of the chapters this dose read.

Only gains attested inside a dose's reading can have been picked up from it. At dose 20 every gain is attested inside its reading (24, 24 and 23 gains in the three runs), and at dose 10, 11 of 12. At doses 1 and 3 most gains are attested only in chapters those doses did not read (6 of 9 and 7 of 11): they move toward her practice without having seen the example, which points to the instruction, the conventions of her English or the reading of her chapters in general rather than to particular renderings.

Where her Chapter 25 repeats her earlier examples, and where it departs

The classification of her Chapter 25 is the instrument's, an unvalidated AI judgment. C0 uses her Chapter 25 option at 103 of the 240 points; the first study's six versions at 90–⁠104.

Use of her Chapter 25 option rises from 104 of 240 points at dose 0 to 121 at dose 20, with single runs of 115, 122, 115 and 114 at doses 1, 3, 10 and 24; the sequence is not monotone. The rise lies where the instrument classifies her Chapter 25 as repeating its earlier examples: 50 to 65 of 78 points. Where it classifies her Chapter 25 as departing from them, every run uses her option at 4–⁠7 of 25 points; where her earlier chapters hold no example, at 25–⁠33 of 85. A low score on the departures can reflect choices tied to this chapter's context, weak analogies in the instrument, missing options, or a failure to imitate; these counts do not distinguish them.

Structure

The structural measures show no clear progression toward her chapter across these runs. Of the 42 coded fragmentary sentences that she completes, the dose versions complete between 9 and 14; of 20 sampled fragments she keeps, every version keeps all of them. At the two passages the German narrates in the present tense she uses the past, and no dose version does; of the 6 tense places coded, every run matches her at the same 2. The refrain keeps the wording the shared reading's plan set out ("What sweetness …", where hers reads "How sweet the air was …"), and the similarity of its returns to one another stays between 0.68 and 0.70 (hers: 0.73). Every dose version keeps the German's 179 paragraph breaks exactly; hers joins two pairs of paragraphs and divides one into three. These are the choices the shared plan had already made (see Design); they persisted in every branch.

Style

Delta (lower is closer). Delta to other machine texts: the mean distance to the other sixteen machine texts in the set (the ten dose runs, C0 and the first study's six). The first study's six machine versions (Claude Opus 4.7): Delta to her Chapter 25 0.903–⁠1.065, to other machine texts 0.500–⁠0.770; em dashes 8.81–⁠18.41 and contractions 9.66–⁠20.56 per 1,000 words. Two of the dose-0 runs and C0 use spaced en dashes in place of em dashes.

The selected word-frequency profile moves toward her Chapter 25 with reading: from 0.903 at dose 0 (runs 0.894–⁠0.908) to 0.813 at dose 20 (runs 0.791–⁠0.847), with doses 1, 3, 10 and 24 between (0.824–⁠0.850). The first study's six versions are 0.903–⁠1.065 from her chapter, C0 0.941. The dose versions stay much nearer the machine versions than hers: 0.43–⁠0.50 on average from the other sixteen, and 0.30–⁠0.48 from one another, against 0.79–⁠0.91 from her Chapter 25. Delta depends on the texts compared and can reflect shared content and source wording as well as style; no reader judged whether any version continues her voice.

The reference branches adopt some conventions of the published English. Every run at doses 1 to 24 uses American spelling and the closed em dash, as her chapters do; at dose 0, two of the three runs use British spelling and spaced en dashes, the third American spelling and the em dash. These conventions may be the publisher's as much as the translator's. The em dash then becomes more frequent than in her text: about 8 per 1,000 words at dose 20 (7.86–⁠8.74) and 6.67 at dose 24, against 3.77 in her Chapter 25 and 2.98 in her chapters 1–⁠24. Contractions rise from about 18 per 1,000 words at dose 0 to 19.66–⁠23.47 at doses 1 to 24; her Chapter 25 has 26.14, her chapters 1–⁠24 have 17.72. Distance to her chapters 1–⁠24, a different text and so larger for every version, does not change with dose: 1.549–⁠1.571 at dose 0, 1.530–⁠1.586 at dose 20, 1.561 at dose 24, against 1.494 for her Chapter 25.

The first study's 45 points

The first study's six versions: 11–⁠18 of the 45 points, chrF 67.3–⁠69.6 (the coding as corrected on 14 October 2026). The 45 points were coded blind in two batches, each with two of the first study's versions as unannounced calibration anchors; all four anchors matched their earlier coding at all 45 points. The dose-0 chrF mean is lowered by one run's quotation marks (Table 6).

On the first study's 45 points, agreement with her Chapter 25 rises from 17 of 45 at dose 0 (runs 16–⁠19) to 22 at dose 20 (runs 20–⁠23); doses 1, 3, 10 and 24 lie between (19–⁠21), and C0 is at 21. chrF rises at dose 1 (70.6 for the first dose-0 run, 71.7 at dose 1) and then stays between 71.4 and 72.2. In the first study the six machine versions agreed with one another at 64% of the 45 points and with her at 31%. By chrF the six first runs, doses 0 to 24, score 86.3–⁠89.8 against one another (mean of both directions).

Run-to-run variation

Run 2 at dose 0 sets its dialogue in single quotation marks, which lowers its chrF; with every quotation mark counted as the same character it scores 70.3, and no other text moves by more than 0.1.

Because the three runs of a dose start from the identical saved reading, their differences show how much a single run of this workflow varies. On the counts, on chrF and on the style measures, the three dose-20 runs lie outside the range of the three dose-0 runs; on fragments they do not differ. Within one dose, two runs put 18 to 34 of the 240 points in different instrument categories. The totals of single runs at doses 1, 3, 10 and 24 (115, 122, 115 and 114 of 240; 101, 104, 105 and 115 of 155) should be read with this variation in mind. Three runs at each end do not establish a dose–response curve.

Since June

The first study's six machine versions were made in June 2026 by Claude Opus 4.7, under six different briefs, from five readings of the novel: D-let and D-aim share one, forked at the moment of translation (see the Method). This study's versions were made in October 2026 by Claude Opus 5.5. One October run, C0, repeats the first study's control arm C on the new model: the same brief and step messages, verbatim, with a reading of its own. C and C0 are the nearest this material comes to the same translation made twice, in June and in October.

C and C0 share a category at 189 of the 240 points (5 of them a shared other) and make the same coded choice at 33 of the first study's 45. C0 sets 6 spaced en dashes in the whole chapter and no em dash.

The single C and C0 outputs have similar aggregate scores, but this comparison changes the model, the independent reading and the runtime around them together, and June's runtime settings are not all recorded; it does not isolate a model effect. Their conventions differ: C sets about 13 em dashes per 1,000 words and spells American, C0 sets none and spells British. Point by point they differ more than their totals suggest: at 51 of the 240 points they fall in different categories, about as often as two of June's differently briefed versions do (Table 9).

Her Chapter 25 is the reference for the first two columns; the third scores it with the same instrument as the machine versions.

Three things differ between the June and October outputs. These comparisons mix model, brief, reading and runtime; they describe the outputs, not the contribution of each.

• Agreement with her Chapter 25 is somewhat higher in October, even without her English: 16–⁠21 of the 45 points against 11–⁠18 in June, and 98–⁠108 of the 240 against 90–⁠104. On the 45 points, June's four versions given a translator's persona agree with her at 11–⁠13, its two controls (one with a permission to domesticate) at 17 and 18. With her chapters, the October branches reach 19–⁠23 of the 45 and 114–⁠124 of the 240.
• The conventions differ. All six June versions set closed em dashes, 8.81–⁠18.41 per 1,000 words. Without her English the October versions set few or none: C0 and two dose-0 runs use spaced en dashes instead, the third dose-0 run sets 4.84 em dashes per 1,000 words. After her chapters every October run sets the closed em dash, 4.37–⁠9.02 per 1,000 words; her Chapter 25 has 3.77. Spelling, three American and three British in June, is mostly British in October until her chapters are read, and American after, as hers is.
• The October versions resemble one another more closely. C0 and the three dose-0 runs share a category at 85–⁠93% of the 240 points, June's six at 71–⁠88%. The dose runs share one reading; C0 has its own and still shares a category with them at 85–⁠88%.

What did not change is what the first study already found. In every version, June or October, 27–⁠32 of the 62 fragments are handled as she handles them, 1–⁠3 of the 6 tense places match hers, and the refrain's returns stay 0.68–⁠0.70 similar to one another, against her 0.73. Where the instrument classifies her Chapter 25 as departing from its earlier examples, every version uses her option at 3–⁠8 of the 25 points.

What the results show

• In this workflow the reference branches use more of her documented earlier renderings. Using an option with a documented earlier rendering rises from 90 at dose 0 to 118 at dose 20, of 155 points (means of three runs). The gains over dose 0 at dose 20 are attested inside the chapters it read; at doses 1 and 3 most are not (Table 1a). The one run given all twenty-four earlier chapters did not improve on the three twenty-chapter runs on these measures.
• Where the instrument classifies her Chapter 25 as repeating its earlier examples, the branches come closer to it (50 to 65 of 78 points). Where it departs from them, or has no earlier example, agreement stays low in every run: 4–⁠7 of 25 and 25–⁠33 of 85 points. The classifications are provisional AI judgments.
• In these branches, several structural choices persisted from the shared reading's plan, and agreement on the instrument's chapter-specific choices remained low. Fragments, tense, the refrain's wording and the paragraphing show no clear progression toward her chapter; every version keeps the German's 179 paragraph breaks.
• The selected word-frequency profile moves toward her Chapter 25 (0.903 to 0.813) and stays far nearer the other machine versions, and the reference branches adopt some conventions of the published English (American spelling, the closed em dash). Delta depends on the texts compared, and no reader evaluated the voice.
• The machine versions remain more similar to one another than to her chapter. On the first study's 45 points the five first runs agree with one another at 81% of the points and with her Chapter 25 at 45%.
• June and October. The single C and C0 outputs have similar aggregate scores but differ in their conventions and in many individual categories; the comparison changes model, reading and runtime together.
• Repeating the machine workflow produced different outputs and different instrument categories. Two runs from the identical saved reading and instructions put 18 to 34 of the 240 points in different categories, and the house style can change from one run to the next (Repeat runs and alternatives).

Repeat runs and alternatives

Translation studies has long asked how much of a translation the source decides. The repeat runs show how much the outputs of one machine workflow differ when nothing is changed but the run; they cannot show how much of her translation was open, or why she chose as she did.

Two versions share a category at a point when the instrument codes both with the same option; each version's choices are those scored above (the script's where the wording decided, the blind coder's elsewhere). A shared other counts as a shared category although it can cover different wordings. The last row is the use of her Chapter 25 option in Table 2. The 240 points are not 240 independent observations: repeated terms and related choices are correlated.

Run again from the identical saved reading, with identical instructions, the workflow puts 18 to 34 of the 240 points in different instrument categories in a second run. At 29 points the three dose-0 runs do not all agree; at 26 of them the script decided all three runs from the wording, so the difference lies in the texts, not in the coding. The least settled layer is the house style: of the three dose-0 runs, two spell British and set spaced en dashes, one spells American and sets closed em dashes, and one sets its dialogue in single quotation marks. Once her chapters have been read, every run takes the published book's conventions.

The splits are not spread evenly. In both sets of three runs they gather in period realia, idioms, dialect and verse (22 and 30 of those 129 points) and are rare in names, titles and foreign words (1 of 47 in each). Process research on human translators finds that their renderings also vary most where the source is hard (Dragsted 2012).

The machine runs often share a category that differs from her Chapter 25: of the 211 points on which the three dose-0 runs agree, her Chapter 25 uses another option at 117. These counts do not determine which rendering the source better supports. At the 29 points where the three runs split, her option is the runs' majority choice at 10, a minority choice at 10 and none of theirs at 9.

What the literature says

Translation theory has long described choices that the source does not fix. Levý (1967) distinguished necessary from unnecessary and motivated from unmotivated decisions, described how the wider context motivates a choice, and observed that translators settle for the solution that gives most effect for least effort. Pym (2003) defined competence as producing more than one viable version and choosing one quickly, with justified confidence. Toury (1995) and Chesterman (1997) turned attention to the norms that make some choices regular in a time and place, while keeping questions of assessment open.

Studies of translations bear this out. Annotators who encoded the acceptable English translations of news sentences in networks found that the networks represent billions of possible translations per sentence, yet reference translations made independently still fell outside them; about half of the remaining differences were optional function words, near-equivalents and spelling variants (Dreyer and Marcu 2012). Five literary translators of one Hemingway story into German resolved its open points differently, according to their stylistic awareness and preferences, knowledge and habitus (Kolb 2013); two German translations of one Fitzgerald novel, published the same year, differ in small recurring choices enough to shift the novel's emphasis (Winters 2007; 2009). These differences are compatible with purposeful interpretation; they say nothing about the reasons behind any particular translator's choices.

For a language model, variation between runs is expected. Translation is uncertain in itself, with many valid translations of a sentence, and translation models spread their probability widely across them (Ott et al. 2018); hosted models set to behave deterministically still give different outputs from run to run (Atıl et al. 2025, on tasks other than translation). This study's literature search found no study that measures run-to-run variation in literary translation by a language model; the search was not exhaustive. The nearest parallel to the convergence measured here is Zhang, Zhao and Eger (2025): in four language pairs, German–English among them, machine translations overlapped with one another and stayed closer to the source's syntax, while the published human translation stood apart.

This study measures agreement, not quality; nothing in it says which choice at a point is better. Whether a difference in quality can be seen at all depends on who judges and how. Judges of machine translation agree only weakly on graded scores, and a judge rescoring the same items agrees with their own earlier scores only moderately (Callison-Burch et al. 2007). Professional translators marking errors with the whole document in view ranked every human translation above every machine system for English–German and Chinese–English news, and still scored three human translations of the same texts clearly differently (Freitag et al. 2021). For literary texts, student raters using an error typology judged nearly 60% of human translations no better than machine output, while best–worst comparison picked out the human translation 80–⁠100% of the time (Zhang, Zhao and Eger 2025). Avid readers preferred published human translations in close reading of aligned passages, yet recognized the human version little better than chance and tended to prefer whichever version they believed was human (Ferstler et al. 2026).

One consequence concerns measurement. A single reference translation is one of many defensible versions, and closeness to it rewards matching its conventions as well as its choices (Freitag, Grangier and Caswell 2020). Here, one dose-0 run scores 67.7 chrF against her chapter, against 70.6 and 70.2 for the other two, only because it sets its dialogue in single quotation marks; with every quotation mark counted as the same character it scores 70.3. The branches that read her chapters better match selected editorial conventions of the published book. Literary quality, and whether reusing a rendering in a new context is an improvement, were not evaluated.

Limits

• One chapter, one model, one reading. All doses are branches of a single reading of the novel by Claude Opus 5.5, compacted once and carrying a developed translation plan. A different reading, or a different order of task (her chapters before any plan), could settle the chapter differently.
• A combined intervention. The reference branches received her chapters, studied them with notes, and were asked to continue in her voice; these parts are not separated. Separating them would need further controls, such as the instruction without her chapters.
• Few repeats. Doses 0 and 20 were run three times; doses 1, 3, 10 and 24, C and C0 once. Differences among single runs of a few points are within the variation between runs. The 240 points are not independent observations.
• An unvalidated instrument. An AI instance built the 240 points, a second one added her chapters 1–⁠4, and a third pass re-examined the evidence at 111 points; the same model family built the instrument, coded the points the script left open, and made the translations. No person has checked the points, their options, or the classification of her practice. The examples mix the same item, comparable items and looser analogies.
• Agreement checks, not accuracy. The checks in How it was measured compare procedures; both can share a mistaken category, and they do not bound how small a difference can be read.
• The 45 points were chosen for the first study's comparison, not for this question.
• Delta and chrF. Delta depends on the set of texts compared; its values are comparable within this page only. chrF counts typographic conventions: quotation marks alone move one run by more than 2.5 points. No reader judged voice, and no assessment of quality or of fidelity to the source was made.
• June and October. C and C0 differ in model, reading and runtime at once, and June's runtime settings are not all recorded.
• Training data. Her translation was published in 2025, and the model may have met it in training. A close reading of the five first-run texts looked for her most distinctive Chapter 25 renderings in the machine versions and found none; this does not rule out exposure.
• The literature. The search for related work (October 2026) covered translation studies, machine-translation research and stylometry. It was not exhaustive.
• A correction in the first study. The first study's published claim about the refrain was corrected on 14 October 2026; see the correction note on the Findings page. Both sets of measurements here use her Chapter 25 with the printed book's paragraphing.
• A working draft. The figures may change with further runs.

Materials

The six first-run translations and the statements their branches wrote, unedited; each file opens with a one-line header. The translations are machine translations of Gabriele Tergit's Chapter 25.

• Dose 0 — translation4,410 words · markdown
• Dose 0 — statement of approach768 words · markdown
• Dose 1 — translation4,341 words · markdown
• Dose 1 — statement of approach765 words · markdown
• Dose 3 — translation4,337 words · markdown
• Dose 3 — statement of approach772 words · markdown
• Dose 10 — translation4,295 words · markdown
• Dose 10 — statement of approach696 words · markdown
• Dose 20 — translation4,300 words · markdown
• Dose 20 — statement of approach768 words · markdown
• Dose 24 — translation4,332 words · markdown
• Dose 24 — statement of approach738 words · markdown

• Results tables (instrument v1.2 and the first study's 45 points)numbers only · markdown

The instrument, its extension, the evidence review and the working analyst's report are not published, because they quote her translation at length; the repeat runs and C0 are not published here. Her translation is published by New York Review Books and is not reproduced here.

References

• Atıl, Berk, et al. 2025. “Non-Determinism of ‘Deterministic’ LLM System Settings in Hosted Environments.” In Proceedings of the 5th Workshop on Evaluation and Comparison of NLP Systems (Eval4NLP), 135–⁠148. ACL. doi.org/10.18653/v1/2025.eval4nlp-1.12
• Callison-Burch, Chris, et al. 2007. “(Meta-) Evaluation of Machine Translation.” In Proceedings of the Second Workshop on Statistical Machine Translation, 136–⁠158. Prague: ACL. aclanthology.org/W07-0718/
• Chesterman, Andrew. 1997. Memes of Translation: The Spread of Ideas in Translation Theory. Amsterdam: John Benjamins. doi.org/10.1075/btl.22
• Dragsted, Barbara. 2012. “Indicators of Difficulty in Translation — Correlating Product and Process Data.” Across Languages and Cultures 13 (1): 81–⁠98. doi.org/10.1556/Acr.13.2012.1.5
• Dreyer, Markus, and Daniel Marcu. 2012. “HyTER: Meaning-Equivalent Semantics for Translation Evaluation.” In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 162–⁠171. Montréal: ACL. aclanthology.org/N12-1017/
• Ferstler, Yves, et al. 2026. “AI Translation of Literary Texts Is ‘Fine’, but Readers Still Prefer Human Translations.” Accepted to Findings of the Association for Computational Linguistics: EMNLP 2026. arXiv:2606.26040. arxiv.org/abs/2606.26040
• Freitag, Markus, David Grangier, and Isaac Caswell. 2020. “BLEU Might Be Guilty but References Are Not Innocent.” In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 61–⁠71. ACL. doi.org/10.18653/v1/2020.emnlp-main.5
• Freitag, Markus, et al. 2021. “Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation.” Transactions of the Association for Computational Linguistics 9: 1460–⁠1474. doi.org/10.1162/tacl_a_00437
• Kolb, Waltraud. 2013. “‘Who Are They?’: Decision-Making in Literary Translation.” In Tracks and Treks in Translation Studies, edited by C. Way, S. Vandepitte, R. Meylaerts and M. Bartłomiejczyk, 207–⁠221. Amsterdam: John Benjamins. doi.org/10.1075/btl.108.11kol
• Levý, Jiří. 1967. “Translation as a Decision Process.” In To Honor Roman Jakobson, vol. 2, 1171–⁠1182. The Hague: Mouton. Reprinted in Scientia Traductionis 11 (2012): 72–⁠96. doi.org/10.5007/1980-4237.2012n11p72
• Ott, Myle, et al. 2018. “Analyzing Uncertainty in Neural Machine Translation.” In Proceedings of the 35th International Conference on Machine Learning, PMLR 80: 3956–⁠3965. proceedings.mlr.press/v80/ott18a.html
• Pym, Anthony. 2003. “Redefining Translation Competence in an Electronic Age: In Defence of a Minimalist Approach.” Meta 48 (4): 481–⁠497. doi.org/10.7202/008533ar
• Toury, Gideon. 1995. Descriptive Translation Studies – and Beyond. Amsterdam: John Benjamins. doi.org/10.1075/btl.4
• Winters, Marion. 2007. “F. Scott Fitzgerald’s Die Schönen und Verdammten: A Corpus-Based Study of Speech-Act Report Verbs as a Feature of Translators’ Style.” Meta 52 (3): 412–⁠425. doi.org/10.7202/016728ar
• Winters, Marion. 2009. “Modal Particles Explained: How Modal Particles Creep into Translations and Reveal Translators’ Styles.” Target 21 (1): 74–⁠97. doi.org/10.1075/target.21.1.04win
• Zhang, Ran, Wei Zhao, and Steffen Eger. 2025. “How Good Are LLMs for Literary Translation, Really? Literary Translation Evaluation with Humans and LLMs.” In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 10961–⁠10988. ACL. doi.org/10.18653/v1/2025.naacl-long.548