Pith. sign in

REVIEW 5 major objections 6 minor 58 references

Mitigating Stylistic Biases of Machine Translation Systems via Monolingual Corpora Only

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Babel is a black-box post-processing framework that detects stylistic inconsistencies in machine translation with 88.21% precision and repairs them using only monolingual corpora, improving style preservation by 150% while keeping…

desk verdict A plausible post-hoc style-repair pipeline whose own Table 3 contradicts the headline 150% improvement; the idea and dataset deserve refereeing, but the paper needs heavy revision. read the letter →

arxiv 2507.13395 v1 pith:42EIC3AO submitted 2025-07-16 cs.CL cs.AI

classification cs.CLcs.AI
keywords machinetranslationstylepreservationdetectiondiffusionmodelsmonolingualcorporablack-boxtestingtransferChinese-English
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces Babel, a black-box post-processing framework that detects and repairs stylistic mismatches in machine translation outputs using only monolingual corpora. It claims that two components, language-specific style detectors and a diffusion-based style applicator, can restore source-text style in translated sentences without parallel stylistic data. Across five domains and six translation systems, Babel reports 88.21% precision in flagging style-inconsistent outputs, an average 35.86% reduction in stylistic bias, a 150% improvement in human-judged style success, and a semantic similarity of 0.92. The practical stake is that style-aware translation could be layered onto any existing system without retraining it or collecting bilingual style data.

What carries the argument

The machinery is a two-module loop. The style detector is a pair of fine-tuned BERT classifiers, one per language, trained on style-annotated monolingual corpora; a style score below the threshold $h = 0.5$ marks a translation as inconsistent. The style applicator is a diffusion model in embedding space, trained with PEGASUS paraphrases as stand-ins for style-neutralized translations, using a slow noise schedule $\beta_t = \sqrt{(T-t)/T}$ to preserve meaning; at inference each denoising step is adjusted by subtracting $\lambda$ times the gradient of the negative cosine similarity between the candidate's style embedding and user-provided style samples, and the best of four candidates is kept only if its semantic similarity to the original translation stays above 0.85. These components together carry the claim that style can be detected and reattached without parallel texts.

What would settle it

Measure the style-score gap between original sentences and their PEGASUS paraphrases and between the same originals and their machine translations in Babel-Corpus; if the paraphrase-induced gap is substantially smaller or differently distributed than the translation-induced gap, the training signal is mismatched to the inference task and the reported repair rates should be re-measured against real translation style loss.

Watch

Extended reading notes

Core claim

The paper establishes that stylistic bias in machine translation can be treated as a test-and-repair problem that does not require parallel corpora. In the testing phase, separate monolingual BERT classifiers score the style of the source and the translated sentence, and a threshold difference flags the translation as style-inconsistent; in the repair phase, a diffusion model trained to reconstruct original sentences from meaning-preserving paraphrases rewrites flagged translations, guided at inference by cosine-similarity gradients toward user-supplied style samples. The central result is that this pipeline finds stylistic inconsistencies with 88.21% precision, lowers the bias ratio across evaluated systems by about 36% on average, and lifts the human-judged style success rate by 150%, from 16% to 34%, while keeping semantic similarity at 0.92. Even the large language models evaluated, GPT-4o and Claude 3.7, show 13.70% and 14.80% bias ratios before repair, comparable to the commercial systems.

Load-bearing premise

The load-bearing premise is that paraphrases produced by PEGASUS keep the meaning but strip away style, so that learning to reconstruct the original from a paraphrase is the same operation as repairing a translation whose style was neutralized.

Editorial extensions

If this is right

  • Babel can be attached to existing translation systems as a black-box post-processor, adding style preservation without retraining the translator or collecting parallel style data.
  • Users can steer repairs toward a target style by supplying example sentences, so style is no longer limited to a preset formal/informal dichotomy.
  • The detector provides a testing signal for translation vendors, flagging style-inconsistent outputs with 88.21% precision and a false positive rate of 10.41%.
  • LLM-based translation also benefits: after Babel, human-judged style success rises from 32% to 54% for GPT-4o and from 32% to 52% for Claude 3.7, with only small semantic losses.
  • Adding a new style domain costs roughly 10.5 hours of fine-tuning in the paper's setup, suggesting new styles can be supported without large parallel resources.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The strongest untested premise is that paraphrase-induced style loss resembles translation-induced style loss; if a language pair or domain violates that, the applicator would be trained on the wrong signal.
  • The modular design suggests transfer to other language pairs by swapping the two BERT detectors and the target-language diffusion backbone, but the paper only demonstrates Chinese-English.
  • The reported 150% improvement starts from a 16% baseline, so the absolute success rate after repair is still only about 34% on average; a deployment would need larger gains to be reliable.
  • Since candidate selection keeps only outputs with semantic similarity above 0.85, Babel is implicitly trading a bounded amount of meaning for style, and the choice of threshold controls that trade-off.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. Babel is a post-processing framework for machine translation that aims to detect and repair stylistic inconsistencies using monolingual, style-labeled corpora. It consists of a language-specific style detector (fine-tuned BERT classifiers) and a diffusion-based style applicator trained to reconstruct original sentences from PEGASUS paraphrases; at inference, the applicator revises flagged MT outputs under user-supplied style guidance and a semantic similarity check. The paper evaluates the framework on Chinese-English translations from six systems across five domains, reporting detection precision of 88.21%, a 150% improvement in stylistic preservation, and an average STS of 0.92, with human ratings and a released dataset/code.

Significance. The paper addresses a real and underexplored problem—stylistic flattening in commercial MT—and its black-box, post-processing formulation is practical. It releases a five-domain bilingual evaluation corpus, reports double-blind human evaluation, and documents computational overhead. The human-evaluation design (Acc/Sem/Flu with a conservative success threshold) is a reasonable way to measure repair quality. If the core claims were supported, Babel would be a useful tool. However, the current manuscript does not substantiate the headline magnitudes, and the automatic evaluation is partly self-referential; the significance will depend on whether the authors can correct the numbers and supply non-circular validation.

major comments (5)
  1. [§4.3, Table 3, and abstract] The headline improvement "from 16% to 34% (+150%)" is contradicted by Table 3's own averages. Averaging the four non-LLM systems gives 16.0% to 33.75%, a relative increase of about 111%; averaging all six systems gives 21.3% to 40.2%, about 89%. The per-system figures in the same paragraph are also off: Youdao is reported as +221% but the table shows +153%, and Opus-MT as +86% but the table shows +73%. The same paragraph's Table 1 reductions (Google 46.87%, Baidu 40.22%, Youdao 31.72%, Opus 24.64%) also differ from the table's own percentages (-46.16%, -42.31%, -32.52%, -22.00%). Because the 150% figure is repeated in the abstract and conclusion, this is not a local typo but a load-bearing error in the paper's central quantitative claim.
  2. [§3.4.3, §4.2, §D.4] The automatic evaluation is circular. Candidate selection in §3.4.3 chooses "the one with the highest style score," and §4.2/D.4 compute the bias ratio and style score improvements using the same style detector. Optimizing against a fixed classifier inflates its own scores even if the underlying stylistic property is unchanged, so Table 1 and the automatic part of §4.3 partly measure the detector's agreement with itself. The Limitations section's caveat that the style score "lacks absolute significance" does not resolve this circularity. Only the human ratings in Table 3 provide independent evidence, and those are the numbers that contain the arithmetic inconsistencies discussed above.
  3. [§3.4.1–3.4.2, Eqs. (5), (6), (11), Algorithm 1] The diffusion formulation is not fully specified and contains internal inconsistencies. Eq. (5) has no dependence on x_{t-1} and reverses the roles of β_t relative to Eq. (1); Eq. (6) defines β_t = sqrt((T-t)/T), which is called a noise schedule but acts as a signal coefficient in Eq. (5). Algorithm 1 line 9 uses ForwardDiffusion(r̂*_0) inside the loop, which should be r̂*_t. The gradient guidance in Eqs. (9)–(10) requires differentiating through discrete top-p sampling, which is not explained. These issues prevent reproduction and make it unclear whether the training objective Eq. (7) matches the inference procedure.
  4. [§3.4.1, §D.1] The central assumption that PEGASUS paraphrases "retain the original meaning but have reduced stylistic elements, simulating the effect of translation" is not validated. At inference the model is applied to actual MT outputs, which may lose style through terminology errors, register shifts, and hallucinations rather than through the neutralization that paraphrasing induces. The authors should measure whether the detector score distributions over paraphrase pairs and over source-translation pairs overlap; if they do not, the training signal is mismatched to the inference task.
  5. [§3.3.1, Appendix A] The style detector is trained on five domain corpora and therefore primarily measures domain membership, not the fine-grained stylistic dimensions (formality, register, rhetorical patterns) claimed in the introduction and used in the qualitative examples. The qualitative repairs in Figure 4 are within-domain register changes (e.g., "根据这里的规定" to "受限于本协议之规定"), but no evidence is given that a domain classifier detects such changes. The paper should either demonstrate within-domain sensitivity of the style score or reframe the detector's role as domain-style matching.
minor comments (6)
  1. [§4.2, Table 2] The text says "the four translation systems" though Table 2 lists six systems; clarify whether the reported aggregate precision and FPR cover the four NMT systems or all six.
  2. [Eqs. (5), (7)] The symbol E(·) denotes both the embedding model in Eq. (5) and the cross-entropy function in Eq. (7); please use separate notation.
  3. [§4.3] The paper states that Fleiss's kappa was calculated but never reports a value; include the inter-annotator agreement statistics for the human evaluation.
  4. [Limitations, §D.3] The placeholder citation "Ano" and the reference "Anonymized Repository - Anonymous GitHub" should be replaced with a stable code/data reference; also, D.3's user customization requires bilingual style samples, which should be reconciled with the "monolingual corpora only" claim.
  5. [Algorithm 1] The variable names in lines 10–14 mix ˆl_t and ˆl*_0 at the final step; the pseudocode should be aligned with Eq. (10).
  6. [§4.3, Table 3] The human evaluation uses only 50 samples per system-domain cell, yet no confidence intervals or significance tests are provided; given the small sample, the aggregate success-rate differences should be reported with uncertainty estimates.

Circularity Check

1 steps flagged · score 5.0 of 10

Automatic style-score and bias-ratio gains in Table 1 are partly circular because the same style detector selects repaired candidates and then supplies the reported style metrics; the headline 150% claim rests on human evaluation and is not circular, though its arithmetic is unsupported by Table 3.

  1. self definitional [Section 3.4.3 (Candidate Selection), Section 4.3 (Experiment Design/Results), Appendix D.4 (Evaluation Metrics), Table 1]
    "For each translation requiring style repair, we generate four candidate outputs and select the one with the highest style score that maintains semantic similarity above 0.85 with the original translation. ... We use a style detector to assess these candidates, considering them repaired if their style matches the original sentence and the semantic loss is within an acceptable threshold. ... These style scores are derived from the confidence provided by the style detector."

    The same style detector is both the selection criterion for keeping a repaired candidate (§3.4.3) and the source of the reported Style Score and Bias Ratio metrics (Appendix D.4, Table 1, §4.3). Selecting the candidate with the highest detector confidence and then reporting detector-confidence gains means the automatic improvements (e.g., 'averaging a 7.20% increase') measure whether the selector increased its own objective; they are not independent evidence of style preservation. The post-repair bias-ratio drop is likewise a filter outcome gated by the same detector. This is partial: the headline 150% figure comes from human ratings in Table 3, not from these automatic metrics, so the central claim retains independent content.

full rationale

The derivation chain of Babel is mostly self-contained: the detector is trained on monolingual style-labeled corpora and its precision (88.21%) is anchored to human labels in Table 2; the applicator is a diffusion model trained to reconstruct originals from paraphrases and then guided by user style samples; no load-bearing uniqueness theorem or self-citation chain is invoked. The single overlapping reference (Zhang et al. 2024b, co-authored by Juan Zhai) appears only in related work and is not load-bearing. The one genuine circular step is the self-scoring repair evaluation: Section 3.4.3 selects the candidate with the highest style score, and Appendix D.4 defines Style Score and Bias Ratio from the same detector's confidence, so the automatic 'style score +7.2%' and 'bias ratio -35.86%' claims in Section 4.3/Table 1 partly reduce by construction to the optimization objective. This does not make the central 150% claim circular because that figure is based on Table 3 human ratings, not the detector; however, those human ratings as reported do not support the 150% magnitude (averaging the six Suc rows gives about 89% relative improvement; Youdao's +221% likewise disagrees with its +153% row). That arithmetic discrepancy is a correctness/support problem rather than circularity and is noted here without increasing the circularity score beyond the partial self-measurement found.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central method rests on strong domain assumptions: style equals domain label, and paraphrase style loss equals translation style loss. The automatic evaluation is circular because the same detector selects and scores candidates. The diffusion noise schedule and inference loop are internally inconsistent, which further weakens the derivation.

free parameters (6)
  • detection threshold h = 0.5
    Tuned in Appendix F to balance precision and false positive rate; used for all detection and repair metrics.
  • inference temperature tau = 0.3
    Tuned in Appendix F to optimize repair effectiveness while controlling semantic loss.
  • guidance strength lambda = 1000
    Tuned in Appendix F; controls the influence of user-supplied style samples on the generated text.
  • number of candidates = 4
    Empirically determined in Section 3.4.3, balancing computational cost and style quality.
  • semantic similarity threshold = 0.85
    Used in candidate selection to ensure revised texts stay close to the original translation.
  • noise schedule beta_t = sqrt((T-t)/T)
    Hand-designed schedule in Eq. 6, chosen to preserve information, but this schedule conflicts with the forward process defined in Eq. 1.
assumptions (5)
  • domain assumption Style can be represented by domain labels and detected by fine-tuned BERT classifiers.
    The paper uses five domains (law, literature, science, medicine, education) as style categories, with no linguistic definition of style beyond domain.
  • ad hoc to paper Paraphrases produced by PEGASUS preserve meaning but strip stylistic elements, simulating translation style loss.
    This is the central training signal for the style applicator (Section 3.4.1); no evidence is provided that paraphrases specifically neutralize style.
  • ad hoc to paper A diffusion model trained to reconstruct a source sentence from its paraphrase can be applied to actual translations at inference.
    The inference procedure feeds the translated text r into the model in place of the paraphrase p (Eq. 8), a transfer that is assumed rather than derived.
  • domain assumption User-supplied style samples and cosine similarity in a learned embedding space provide cross-lingual style guidance.
    The style guidance term J in Eq. 9 relies on a style embedding model Es that is never specified, so the cross-lingual alignment of styles is assumed.
  • standard math Standard diffusion model math (Eqs. 1-3) provides the background for the method.
    The paper invokes standard diffusion definitions, though the subsequent equations are inconsistent with them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Stylistic Biases of Machine Translation Systems via Monolingual Corpora Only." pith.science (2026). https://pith.science/paper/42EIC3AO

@misc{pith2026250713395,
  author       = {Pith},
  title        = {Pith review of: Mitigating Stylistic Biases of Machine Translation Systems via Monolingual Corpora Only},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42EIC3AO}},
  note         = {Machine review of arXiv:2507.13395}
}
read the original abstract

The advent of neural machine translation (NMT) has revolutionized cross-lingual communication, yet preserving stylistic nuances remains a significant challenge. While existing approaches often require parallel corpora for style preservation, we introduce Babel, a novel framework that enhances stylistic fidelity in NMT using only monolingual corpora. Babel employs two key components: (1) a style detector based on contextual embeddings that identifies stylistic disparities between source and target texts, and (2) a diffusion-based style applicator that rectifies stylistic inconsistencies while maintaining semantic integrity. Our framework integrates with existing NMT systems as a post-processing module, enabling style-aware translation without requiring architectural modifications or parallel stylistic data. Extensive experiments on five diverse domains (law, literature, scientific writing, medicine, and educational content) demonstrate Babel's effectiveness: it identifies stylistic inconsistencies with 88.21% precision and improves stylistic preservation by 150% while maintaining a high semantic similarity score of 0.92. Human evaluation confirms that translations refined by Babel better preserve source text style while maintaining fluency and adequacy.

Figures

Figures reproduced from arXiv: 2507.13395 by the authors.

Figure 1
Figure 1. Overview of Babel. 3.3 Style Detector The style detector determines the style at￾tributes of the source text and evaluates whether these attributes are maintained in the translation. The primary challenge is accu￾rately identifying and matching stylistic fea￾tures across different languages, as each lan￾guage has distinct stylistic norms and expres￾sions. To address this challenge, we train a model to recognize and … view at source ↗
Figure 2
Figure 2. Effect of h on the average performance of Babel’s testing process. tially depending on the size of the datasets and the specific languages involved. F Impact of Configurable Parameters Experiment Design: Babel leverages three hyperparameters: detection threshold h, infer￾ence temperature τ , and guidance strength λ, to find and repair stylistic consistent issues. The detection threshold h determines the point at whi… view at source ↗
Figure 3
Figure 3. Effect of τ and λ on the average performance of Babel’s repairing process. able false positive rate, suggesting an optimal balance at h = 0.5. For repair performance, the parameter τ shows that allowing moder￾ate lexical deviations (τ = 0.3) optimizes the number of corrected stylistic issues, while a higher τ value can detrimentally affect seman￾tic integrity. The guidance strength parameter λ demonstrates that mode… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Example of stylistic inconsistent issues and repaired translation generated by Babel. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 38 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    https://huggingface.co/datasets/ACOSharma/literature

    ACOSharma /literature Datasets at Hugging Face . https://huggingface.co/datasets/ACOSharma/literature

  4. [4]

    https://huggingface.co/datasets/AlekseyKorshuk/fairy-tale-books

    AlekseyKorshuk /fairy-tale-books Datasets at Hugging Face . https://huggingface.co/datasets/AlekseyKorshuk/fairy-tale-books

  5. [5]

    https://anonymous.4open.science/r/Babel-3EB2/README.md

    Anonymized Repository - Anonymous GitHub . https://anonymous.4open.science/r/Babel-3EB2/README.md

  6. [6]

    https://huggingface.co/datasets/twang2218/chinese-law-and-regulations

    Twang2218/chinese-law-and-regulations Datasets at Hugging Face . https://huggingface.co/datasets/twang2218/chinese-law-and-regulations

  7. [7]

    https://www.chinahighlights.com/travelguide/english-levels-in-china.htm#/ English levels in china: Quality of spoken english, signage, etc

    2020. https://www.chinahighlights.com/travelguide/english-levels-in-china.htm#/ English levels in china: Quality of spoken english, signage, etc

  8. [8]

    Sentence-transformers/all- MiniLM-L6-v2 Hugging Face

    2024. Sentence-transformers/all- MiniLM-L6-v2 Hugging Face . https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2

Show all 58 references
  1. [9]

    https://www.anthropic.com/news/claude-3-7-sonnet Claude 3.7 Sonnet and Claude Code

    2025. https://www.anthropic.com/news/claude-3-7-sonnet Claude 3.7 Sonnet and Claude Code

  2. [10]

    https://openai.com/index/hello-gpt-4o/ Hello GPT-4o

    2025. https://openai.com/index/hello-gpt-4o/ Hello GPT-4o

  3. [11]

    American Psychological Association. 2019. https://faculty.tamuc.edu/jdavis/tmgt/599/242/TMGT599-242-Syllabus.pdf Publication manual of the american psychological association,(2020) . 428

  4. [12]

    Eugene Bagdasaryan and Vitaly Shmatikov. 2022. Spinning language models: Risks of propaganda-as-a-service and countermeasures. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1532--1532

  5. [13]

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulnar Erlingsson, et al. 2021. Extracting training data from large language models. In USENIX Security Symposium, pages 2633--2650

  6. [14]

    Dhivya Chandrasekaran and Vijay Mago. 2022. https://doi.org/10.1145/3440755 Evolution of Semantic Similarity -- A Survey . ACM Computing Surveys, 54(2):1--37

  7. [15]

    Jianlin Chen. 2024. https://doi.org/10.48550/arXiv.2403.08943 LMStyle Benchmark : Evaluating Text Style Transfer for Chatbots . Preprint, arXiv:2403.08943

  8. [16]

    Simin Chen, Cong Liu, Mirazul Haque, Zihe Song, and Wei Yang. 2022. Nmtsloth: understanding and testing efficiency degradation of neural machine translation systems. In Proceedings of the 30th ACM Joint European Software Engineering Conference and Symposium on the Foundations ...

  9. [17]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://doi.org/10.48550/arXiv.1911.02116 Unsupervised Cross-lingual Representation Learning at Scale...

  10. [18]

    Yiming Cui, Ting Liu, Zhipeng Chen, Shijin Wang, and Guoping Hu. 2016. Consensus attention-based neural networks for chinese reading comprehension. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1777--1786...

  11. [19]

    Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. 2019. Style transformer: Unpaired text style transfer without disentangled latent representation. arXiv preprint arXiv:1905.05621

  12. [20]

    Daryna Dementieva, Daniil Moskovskiy, David Dale, and Alexander Panchenko. 2023. https://doi.org/10.18653/v1/2023.ijcnlp-main.70 Exploring Methods for Cross-lingual Text Style Transfer : The Case of Text Detoxification . In Proceedings of the 13th International Joint Conferenc...

  13. [21]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. https://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding

  14. [22]

    Joshua A Fishman. 2020. Who speaks what language to whom and when? In The bilingualism reader, pages 55--70. Routledge

  15. [23]

    Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Proceedings of the AAAI conference on artificial intelligence, volume 32

  16. [24]

    Shashij Gupta, Pinjia He, Clara Meister, and Zhendong Su. 2020. Machine translation testing via pathological invariance. In Proceedings of the 28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 863--875

  17. [25]

    Xiaochuang Han, Sachin Kumar, and Yulia Tsvetkov. 2023. https://doi.org/10.48550/arXiv.2210.17432 SSD-LM : Semi-autoregressive Simplex-based Diffusion Language Model for Text Generation and Modular Control . Preprint, arxiv:2210.17432

  18. [26]

    Pinjia He, Clara Meister, and Zhendong Su. 2020. Structure-invariant testing for machine translation. In Proceedings of the ACM/IEEE 42nd International Conference on Software Engineering, pages 961--973

  19. [27]

    Pinjia He, Clara Meister, and Zhendong Su. 2021. Testing machine translation via referential transparency. In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pages 410--422. IEEE

  20. [28]

    Georg Heigold, G \"u nter Neumann, and Josef van Genabith. 2017. How robust are character-based word embeddings in tagging and mt against wrod scramlbing or randdm nouse? arXiv preprint arXiv:1704.04441

  21. [29]

    Zachary Horvitz, Ajay Patel, Chris Callison-Burch, Zhou Yu, and Kathleen McKeown. 2024. https://doi.org/10.48550/arXiv.2308.15459 ParaGuide : Guided Diffusion Paraphrasers for Plug-and-Play Textual Style Transfer . Preprint, arxiv:2308.15459

  22. [30]

    You Sound Just Like Your Father

    Dirk Hovy, Federico Bianchi, and Tommaso Fornaciari. 2020. https://doi.org/10.18653/v1/2020.acl-main.154 “ You Sound Just Like Your Father ” Commercial Machine Translation Systems Include Stylistic Biases . In Proceedings of the 58th Annual Meeting of the Association for Compu...

  23. [31]

    Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In International conference on machine learning, pages 1587--1596. PMLR

  24. [32]

    Pin Ji, Yang Feng, Jia Liu, Zhihong Zhao, and Baowen Xu. 2021. Automated testing for machine translation via constituency invariance. In 2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE), pages 468--479. IEEE

  25. [33]

    Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. https://doi.org/10.1162/coli_a_00426 Deep Learning for Text Style Transfer : A Survey . 48(1):155--205

  26. [34]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. 2019. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on...

  27. [35]

    Dongyeop Kang and Eduard Hovy. 2021. https://doi.org/10.18653/v1/2021.acl-long.185 Style is NOT a single variable: Case Studies for Cross-Stylistic Language Understanding . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th ...

  28. [36]

    Jonathan Li, Rohan Bhambhoria, and Xiaodan Zhu. 2022 a . https://aclanthology.org/2022.nllp-1.10 Parameter-efficient legal domain adaptation . In Proceedings of the Natural Legal Language Processing Workshop 2022, pages 119--129, Abu Dhabi, United Arab Emirates (Hybrid). Assoc...

  29. [37]

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. 2022 b . Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328--4343

  30. [38]

    Qingyi Liu, Jinghui Qin, Wenxuan Ye, Hao Mou, Yuxuan He, and Keze Wang. 2024. https://doi.org/10.1609/aaai.v38i17.29832 Adaptive prompt routing for arbitrary text style transfer with pre-trained language models . In Proceedings of the Thirty-Eighth AAAI Conference on Artificia...

  31. [39]

    Guoqing Luo, Yu Han, Lili Mou, and Mauajama Firdaus. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.381 Prompt- Based Editing for Text Style Transfer . In Findings of the Association for Computational Linguistics : EMNLP 2023 , pages 5740--5750. Association for Computat...

  32. [40]

    Huiyu Mai, Wenhao Jiang, and Zhi-Hong Deng. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.990 Prefix- Tuning Based Unsupervised Text Style Transfer . In Findings of the Association for Computational Linguistics : EMNLP 2023 , pages 14847--14856. Association for Computa...

  33. [41]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. https://arxiv.org/abs/1609.07843 Pointer sentinel mixture models . Preprint, arXiv:1609.07843

  34. [42]

    Ojha, John P

    Sourabrata Mukherjee, Akanksha Bansal, Atul Kr. Ojha, John P. McCrae, and Ondrej Dusek. 2023. https://aclanthology.org/2023.icon-1.13/ Text Detoxification as Style Transfer in English and Hindi . In Proceedings of the 20th International Conference on Natural Language Processin...

  35. [43]

    Lei Pan, Yunshi Lan, Yang Li, and Weining Qian. 2024. https://doi.org/10.48550/arXiv.2402.13647 Unsupervised Text Style Transfer via LLMs and Attention Masking with Multi-way Interactions . Preprint, arXiv:2402.13647

  36. [44]

    Daniel Pesu, Zhi Quan Zhou, Jingfeng Zhen, and Dave Towey. 2018. A monte carlo method for metamorphic testing of machine translation services. In Proceedings of the 3rd International Workshop on Metamorphic Testing, pages 38--45

  37. [45]

    Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. Advances in neural information processing systems, 30

  38. [46]

    Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics, pages 4275--4293

  39. [47]

    Zeyu Sun, Zhenpeng Chen, Jie Zhang, and Dan Hao. 2024. Fairness testing of machine translation systems. ACM Transactions on Software Engineering and Methodology

  40. [48]

    Zeyu Sun, Jie M Zhang, Mark Harman, Mike Papadakis, and Lu Zhang. 2020. Automatic testing and improvement of machine translation. In Proceedings of the ACM/IEEE 42nd international conference on software engineering, pages 974--985

  41. [49]

    Zeyu Sun, Jie M Zhang, Yingfei Xiong, Mark Harman, Mike Papadakis, and Lu Zhang. 2022. Improving machine translation systems via isotopic replacement. In Proceedings of the 44th international conference on software engineering, pages 1181--1192

  42. [50]

    J \"o rg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT — B uilding open translation services for the W orld. In Proceedings of the 22nd Annual Conferenec of the European Association for Machine Translation (EAMT), Lisbon, Portugal

  43. [51]

    Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, et al. 2022. Taxonomy of risks posed by language models. 2022 ACM Conference on Fairness, Accountability, and Transparency, pag...

  44. [52]

    Bright Xu. 2019. https://doi.org/10.5281/zenodo.3402023 Nlp chinese corpus: Large scale chinese corpus for nlp

  45. [53]

    Jingjing Xu, Ji Wen, Xu Sun, and Qi Su. 2017. A discourse-level named entity recognition and relation extraction dataset for chinese literature text. volume abs/1711.07010

  46. [54]

    Chiyu Zhang, Honglong Cai, Yuezhang Li, Yuexin Wu, Le Hou, and Muhammad Abdul-Mageed. 2024 a . https://doi.org/10.18653/v1/2024.naacl-srw.21 Distilling Text Style Transfer With Self-Explanation From LLMs . In Proceedings of the 2024 Conference of the North American Chapter of ...

  47. [55]

    Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In International conference on machine learning, pages 11328--11339. PMLR

  48. [56]

    Ningyu Zhang, Mosha Chen, Zhen Bi, Xiaozhuan Liang, Lei Li, Xin Shang, Kangping Yin, Chuanqi Tan, Jian Xu, Fei Huang, Luo Si, Yuan Ni, Guotong Xie, Zhifang Sui, Baobao Chang, Hui Zong, Zheng Yuan, Linfeng Li, Jun Yan, Hongying Zan, Kunli Zhang, Buzhou Tang, and Qingcai Chen. 2...

  49. [57]

    Quanjun Zhang, Juan Zhai, Chunrong Fang, Jiawei Liu, Weisong Sun, Haichuan Hu, and Qingyu Wang. 2024 b . Machine translation testing via syntactic tree pruning. ACM Transactions on Software Engineering and Methodology

  50. [58]

    Zhirui Zhang, Shuo Ren, Shujie Liu, Jianyong Wang, Peng Chen, Mu Li, Ming Zhou, and Enhong Chen. 2018. Style transfer as unsupervised machine translation. arXiv preprint arXiv:1808.07894

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.