Creativity in translation: machine translation as a constraint for literary texts

Introduction

• The earlier experiment used a text emphasizing action and emotion, tested only English-to-Catalan, and relied on one professional translator as judge.The subsequent experiment adjusts the methodology and includes interviews with nine professional literary translators.Source: Introduction · p. 2

Related Work

• Translation-creativity research describes creative solutions as novel departures that preserve meaning and fit textual, situational, and cultural contexts.The reviewed framework also treats creativity as involving acceptability, flexibility, novelty, and fluency.Source: Related Work · p. 3
• Recent literary-translation research examines MT usability, its effects on translators and students, and professionals’ opinions of technology.The article frames its study as part of this research landscape.Source: Introduction · p. 1, Related Work · p. 3
• In the reported results, only 23% of sentences for GT and 21% for DeepL were found to be without errors, giving the error-free rates reported for both systems.The authors state that genre and authorial style play a fundamental role when translating literature because literary texts are less homogeneous than technical texts.Source: Related Work · p. 4
• Studies of post-editing report priming, reduced creativity, or diminished translator voice, although they do not uniformly find more errors than in human translation.Reported post-editing studies cover English-to-Turkish, German, and Slovene.Source: Related Work · p. 4
• Surveys find literary translators use general digital resources but are less inclined toward CAT tools or MT, citing suitability, creativity, and lexical and syntactic variation.Reported reluctance may also relate to familiarity with developments, training, or age.Source: Related Work · p. 4

Research Methodology

• The first CREAMT axis examines creativity across translation modalities; a separate axis concerns reader engagement, enjoyment, and reception.The reported study addresses the first axis and asks whether creativity can be quantified and whether it differs by modality.Source: Research Methodology · p. 4
• Because the study had four translators and five reviewers, the researchers examined available data from multiple angles, including both translated texts and the translation process.They describe these complementary data points as supporting a richer view of MT’s effects on creativity.Source: Research Methodology · p. 5

Source Text

• The researchers selected Vonnegut’s 2BR02B to meet criteria including creative potential, engagement, short length, and accessibility across generations.They also required a text absent from the MT engine’s training data; the story concerns a future where each birth requires another person’s death.Source: Source Text · p. 5
• The story has 123 paragraphs, 234 segments, and 2,548 words; the authors report it had not been translated into Catalan or Dutch to their knowledge.Its reading-ease measures indicate it is fairly easy for an English speaker to read.Source: Source Text · p. 5

Machine Translation

• The English-to-Catalan MT system used literary and target-language data alongside more than four million sentence pairs; the English-to-Dutch system used about five million sentence pairs from parallel novels.Neither system’s training data included the experimental story or any text by Vonnegut.Source: Machine Translation · p. 6

Translation Process

• Four professional literary translators produced HT and PE versions, using a paragraph-segmented tool and a cross-over arrangement between translating and post-editing halves.When combining translators’ work, the researchers resolved inconsistencies based on frequency or order of appearance.Source: Translation Process · p. 6
• Catalan translators rated the MT engine more favorably than Dutch translators, and translators reported faster work but higher cognitive and technical effort in PE.T3 and T4 said they lacked PE experience; T4 described the engine’s proposals as poor and preferred translating to post-editing.Source: Translation Process · p. 7

Review process

• The five reviewers had five or more years of experience and had translated between 15 and 40 novels individually or as part of a team.Only one reviewer reported experience in post-editing.Source: Review process · p. 7
• Reviewers assessed three coded texts while blind to modality: A was PE, B was MT, and C was HT.They received instructions and forms for their assessments.Source: Review process · p. 7
• After reviewing the texts, reviewers completed a short questionnaire about overall translation quality.The reviewers were professional literary translators.Source: Review process · p. 7

Measuring Creativity

• Creativity combines acceptability (value and fitness for purpose) with novelty (newness and originality), rather than measuring novelty alone.The authors operationalize acceptability through error counts and severity plus Kudos, and novelty through creative solutions for source-text problems.Source: Measuring Creativity · p. 8
• Two experienced researchers and translators marked source-text units expected to require problem-solving, aggregating their annotations into 185 units across 123 paragraphs.They agreed on 77 units and identified 108 different units; the annotations were aggregated because problems could vary by language and annotator.Source: Measuring Creativity · pp. 8–9
• Reviewers classified target-text solutions for each unit as reproduction, omission, error, or creative shift, with shifts including abstraction, concretisation, and modification.The classification scheme also distinguishes modification subtypes and reproductions such as retention or direct translation.Source: Measuring Creativity · p. 9
• The creativity score combines creative shifts, unit coverage, errors, and Kudos, then multiplies the result by 100 for a percentage.The equation normalizes by the number of words in the source text.Source: Measuring Creativity · pp. 8, 10

Results

• The results section is divided into translation-process findings, review findings, and interview analysis.Source: Results · p. 10
• One part reports results from the translation process.Source: Results · p. 10
• The remaining parts report review results and interview analysis.Source: Results · p. 10

Results on translators’ changes and effort

• HT was further from the MT output than PE in both languages: CA means were 61.96 versus 54.72, and NL means were 58.81 versus 41.82.Figure 1 presents HTER by language and modality.Source: Figure 1 · p. 11, Results on translators’ changes and… · p. 11
• CA translators made more PE changes than NL translators, while T4’s PE stayed closest to the MT proposal.Both NL translators rated MT output lowest but made fewer PE changes; the authors suggest unfamiliarity with PE may have encouraged preserving more NMT output.Source: Results on translators’ changes and… · p. 12
• The authors report that the output was not suitable for production and that CA PE HTER exceeded the previous study’s result despite a customized, improved engine.They suggest the newer source text’s greater creativity may have required more changes for optimal final quality.Source: Results on translators’ changes and… · p. 12

Translation effort according to PET

• Average edit time was close between HT (216.36) and PE (219.80), while NL took longer than CA in both activities.Means were CA HT 196.42 and NL HT 236.22; CA PE 165.81 and NL PE 273.98.Source: Translation effort according to PET · pp. 12–13
• In this small sample, average keystrokes were much higher for HT than PE (194.03 versus 90.98); the authors caution that general conclusions cannot be drawn.Source: Translation effort according to PET · p. 13
• Pauses were more frequent in HT than PE (13.36 versus 10.12 on average), which the authors interpret as suggesting greater cognitive demand during translation.They hypothesize that pauses may reflect time spent searching for more creative solutions.Source: Translation effort according to PET · p. 13
• EN-CA translators revisited segments more often, potentially indicating more revision, but time spent in the tool may not capture incubation outside it.The authors also note that the sample is small, so general conclusions cannot be drawn.Source: Translation effort according to PET · pp. 13–14

• Reviewers unanimously rated HT Extremely good, MT Extremely bad, and PE Neither good nor bad; CA PE received a higher rating than NL PE.Source: Results from the review · p. 14
• PE had double HT’s errors in CA and three times HT’s errors in NL, despite the same translators working on both modalities.The authors suggest the text’s high number of UCPs could be a reason.Source: Table 3 · p. 15, Results from the review · p. 15
• Interviewed translators said PE limited their creativity and was technically and cognitively more demanding.Source: Results from the review · p. 15
• Fleiss kappa was below 0.40 in five modality-language cases, indicating poor agreement, and was 0.50 for NL MT; reviewers agreed on overall quality more than on individual errors.Neutral errors were included in the calculation.Source: Results from the review · pp. 15–16
• HT had the most creative shifts, followed by PE and MT; HT was especially ahead of PE in Modification, adapting the source text to target cultures.The reviewers classified reproductions and creative shifts in 185 UCPs.Source: Table 7 · p. 18, Results from the review · p. 18
• HT exceeded PE by 10 points in CA and 16 in NL, and exceeded MT by 51 points in CA and 48 in NL on creativity scores.The reported ranking was HT, then PE, then MT; using MT in PE also resulted in a less creative literary translation.Source: Table 10 · p. 20, Results from the review · p. 20

5. The ideal tool for literary translation is human centred

• Participants saw MT as potentially useful for mechanical or straightforward passages, but said it misses context, style, and whole-text coherence.They preferred a human-controlled tool that supports translators rather than supplying a translation proposal; some reported using MT ideas in moderation.Source: §1 · p. 21, §5 · pp. 23–24

2. Creativity is a problem-solving process

• Creativity involves finding surprising solutions that sound natural in the target text while fitting the author’s context and style.Participants also described cultural mediation and imagining the world depicted by the author as part of the work.Source: §2 · pp. 21–22

3. The proposals act as a constraint to creativity

• Post-editing proposals can condition translators, make it harder to move beyond the first solution, and shift attention away from whole-text coherence and style.Translators described PE as requiring greater cognitive, technical, and creative effort than translating independently.Source: §3 · p. 22

4. The delicate equilibrium of reviewing

• Reviewers praised the HT text, placed PE between HT and MT as less creative, and said MT required substantial intervention to become readable.Reviewers were surprised to learn that the same translators had worked in HT and PE, while they sometimes found MT more fluent than expected.Source: §4 · pp. 22–23

6. The technology or modality impact depends on the type of reader

• Readers’ experiences may differ: some accept imperfect translations when focused on plot, while others notice errors and inconsistencies.Participants expected HT to read more naturally and be more entertaining; they described MT as challenging for both reader types.Source: §6 · p. 24

7. The final translation as a product of many collaborative steps

• Literary translation was described as a collaborative, iterative process involving multiple drafts, reviews, and feedback from editors, native speakers, and others.Review practices vary by translator and text, with some passages receiving intensive sentence-level work.Source: §7 · pp. 24–25

8. The grim future of translation and technology

• Four participants voiced worry that improving MT could lead publishers to reduce already low literary translation rates or eliminate translators.Participants also anticipated that MT would improve over time.Source: §8 · p. 25

Conclusions

• HT scored higher for creativity than both PE and MT; professional translators intervened in both highest-scoring modalities.The experiment confirms the authors’ previous findings.Source: Conclusions · p. 25
• Pauses seemed to accompany effortful creative problem-solving that translators enjoyed, so cognitive effort did not necessarily indicate an unhappy process.Source: Conclusions · p. 25
• Across English–Catalan and English–Dutch, creative-shift analysis found HT more novel and flexible, less constrained by MT, and associated with fewer errors.Different expert reviewers and interviews also confirmed the previous results.Source: Conclusions · p. 26
• During PE, translators were less creative, with more errors and fewer creative shifts, even when the same translators worked in both modalities.Source: Conclusions · p. 26
• The proposed definition describes creative translation as identifying a source problem, generating novel and elegant solutions, then choosing one suited to the target text and culture.Source: Conclusions · p. 26
• The authors link PE’s constraint to translators evaluating MT output before forming their own solution; they do not know whether offering MT after an initial translation would differ.Source: Conclusions · p. 26

Funding information

• The project received funding from the European Union’s Horizon 2020 research and innovation programme under Marie Skłodowska-Curie grant agreement No. 890697.Source: Funding information · p. 26