Pith. sign in

REVIEW 4 major objections 5 minor 60 references

Towards Style Alignment in Cross-Cultural Translation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that LLMs flatten style in cross-cultural translation, and that retrieving native exemplars via learned embedding directions restores style alignment by up to 56% on GPT-4 without degrading translation quality.

desk verdict Solid characterization of LLM style-translation failures; the RASTA improvement claim is plausible but the headline numbers are partly circular and the human study too weak to settle it. read the letter →

arxiv 2507.00216 v1 pith:G3VNJMAL submitted 2025-06-30 cs.CL

Shreya Havaldar , Adam Stein , Eric Wong , Lyle Ungar This is my paper
classification cs.CL
keywords stylealignmentcross-culturaltranslationretrieval-augmentedgenerationpolitenesspreservationformalityintimacyAnglocentricbiasLLMevaluation
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

The paper argues that successful cross-cultural translation requires aligning the style a speaker intends with the style a listener perceives, not just transferring literal meaning. It claims that today's LLMs systematically fail this task: they flatten translated style toward neutrality, perform worse when translating into non-Western languages, and produce outputs that standard translation-quality metrics do not flag. To fix the failure, it proposes RASTA, which computes an embedding-space direction that separates translated text from native text for a given style and language pair, then uses that direction to retrieve native exemplars as few-shot prompts. On GPT-4, RASTA raises the style alignment score $A$, defined as the correlation between a per-language style quantifier's score on the original text and its score on the translated text, by up to 56% across politeness, intimacy, and formality, while leaving GEMBA and COMET KIWI scores essentially unchanged.

What carries the argument

The load-bearing object is the cultural alignment mapping $v_{\mathrm{align}}(L_1,L_2,S)=v_{\mathrm{native}}-v_{\mathrm{trans}}$, where $v_{\mathrm{native}}=\mu(L_2,S)-\mu(L_1,S)$ is the embedding-space shift between native text of style $S$ in the source and target languages and $v_{\mathrm{trans}}=\mu(L_1\to L_2,S)-\mu(L_1,S)$ is the shift that translation itself introduces. Adding $v_{\mathrm{align}}$ to an input embedding moves it to the region where native target-language text of style $S$ lies, and the five training-set sentences nearest to that moved point are fed to the LLM as few-shot exemplars. The per-language Mistral-7B style quantifiers supply the scores whose correlation defines $A$, and the BGE-M3 embedding model supplies the geometry in which the centroids and mappings are computed.

What would settle it

Have native bilingual annotators rate the perceived style of the original and RASTA-translated sentences on a continuous scale; if the correlation between those ratings is not higher for RASTA than for the "preserve style" prompt, the claimed alignment improvement is an artifact of the quantifier-based metric rather than a real gain in perceived style.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that style alignment in LLM translation is measurable as the correlation $A$ between a per-language style quantifier's score on the source text and its score on the translation, and that this quantity is currently low, especially when translating into Japanese, Chinese, and Brazilian Portuguese. It further establishes that LLMs shrink the variance of style in translated text relative to native text, and that established reference-free metrics like GEMBA and COMET KIWI correlate negatively or not at all with $A$. The central positive result is that RASTA, by retrieving target-language exemplars whose embeddings lie near a style-aligned target point, raises $A$ substantially on GPT-4, with average relative improvements of 32.1% for politeness, 22.2% for intimacy, and 56.3% for formality, while GEMBA and COMET KIWI degrade by under 1.5%. Bilingual annotators preferred RASTA over an explicit "preserve style" prompt 61% of the time for politeness and 63% for formality.

Load-bearing premise

The chain of measurement assumes that the separately trained per-language style quantifiers put different languages on the same style scale, so that a 0.7 politeness score in Japanese means the same degree of politeness as 0.7 in English, and that correlation between those scores faithfully matches how a human listener would perceive style.

Editorial extensions

If this is right

  • Style alignment should be tracked as its own evaluation axis, because GEMBA and COMET KIWI correlate negatively or not at all with $A$ and therefore cannot catch style flattening.
  • Retrieval-augmented prompting with native exemplars is a viable, training-free way to restore style in LLM translation, at least for high-resource languages and the three styles tested.
  • The Anglocentric gap in style alignment can be narrowed: for GPT-4 politeness, RASTA reduces the gap between Western and non-Western target languages from 0.35 to 0.12.
  • RASTA counteracts the neutrality bias by raising the standard deviation of translated politeness scores toward native levels, a 36% average increase on GPT-4.

Reading between the lines

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

  • A control the paper does not run: replace $v_{\mathrm{align}}$ with a random vector of the same norm during exemplar retrieval; if $A$ does not drop, the improvement comes from having in-context examples rather than from the learned direction.
  • Because RASTA only needs native target-language text with style labels, it could in principle extend to low-resource languages, but the paper tests only high-resource languages, so that extension remains untested.
  • The strongest check on the metric would be a human continuous-rating study of perceived style for both source and translated sentences; the paper's binary preference study is suggestive but does not directly validate $A$.
  • The neutrality-bias finding implies a cheap diagnostic for any translation system: compare the style-score distribution of its output with native text in the target language, not just pairwise correlations.
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

4 major / 5 minor

Summary. The paper defines style alignment as the correlation between intended and interpreted style scores (Eq. (2)), trains per-language Mistral-7B regressors for politeness, intimacy, and formality, and uses them to show that LLM translations are biased toward neutral style, are worse for non-Western target languages, and are not captured by GEMBA/COMET. It then proposes RASTA, which uses differences between native and translated embedding centroids to retrieve few-shot exemplars and prompt the LLM to match a target style level. The paper reports up to 56% relative improvement in A with GPT-4 without large degradation in GEMBA/COMET, and a modest human preference for RASTA over a 'preserve-style' prompt baseline.

Significance. If validated, the paper opens a practical direction: style alignment can be improved by retrieval-augmented prompting without model fine-tuning. The characterization experiments are mostly clear and the neutrality-bias observation is well supported. The authors make code, data, and prompts available, and the human validation study is an honest attempt at external grounding. However, the central quantitative claim currently rests on a metric that is built from the same style regressors that are used to construct the RASTA prompt, and the independent human evidence is too weak to fully certify the claimed effect sizes. The paper is likely to be influential for cross-cultural MT evaluation, provided the measurement concerns are addressed.

major comments (4)
  1. [Section 2, Eq. (2)] The style alignment metric A is a Pearson correlation, but the paper's definition of perfect alignment (Eq. (1)) requires equality of scores, C1(x) = C2(T(x)). Correlation does not penalize systematic score offsets or scaling differences between the separately trained per-language quantifiers; a translation model that consistently shifts style by a constant amount could still attain high A. Since the paper's central failure analysis (Figs. 3-5) and its RASTA improvements all use A, the authors should show that the quantifier scores are cross-linguistically comparable (e.g., through calibrated score distributions or by reporting an agreement metric such as MAE alongside A). Without this, the reported alignment gains may overstate how well interpreted style matches intended style.
  2. [Section 4.3 / Figure A7 and Eq. (2)] There is a potential circularity: the RASTA prompt explicitly provides a numeric style level for the source text and instructs the model to match it, while A is computed using quantifiers trained with the same recipe and on the same datasets. The LLM can therefore increase A by producing text that the target-language quantifier scores near the provided value, without necessarily changing the style perceived by human listeners. The Prolific study (Section 5.2, Appendix F) is the only external check, but it uses 18 annotators, 30 items per language, has pairwise agreement of only 0.167-0.247, and compares RASTA only against the 'preserve-style' baseline. This is too weak to certify the large improvements in Table 3. I recommend adding a validation with a different style classifier, or a human rating study with more annotators and an explicit test of perceived style levels (not just pairwise preference).
  3. [Section 5.1 / Table 3 and Table A6] The paper reports relative gains in A up to 56% but gives no confidence intervals or significance tests for any of the Table 3 comparisons. Because A is a correlation over test corpora, bootstrap intervals or permutation tests are needed to establish that the gains are not noise, particularly for language pairs with small test sets (e.g., Chinese intimacy, 0.28 to 0.39). In addition, the main-text claim of 'no significant degradation' is stated without a model qualifier; Table A6 shows that RASTA degrades GEMBA by 21-27% for Llama-3.2 on intimacy and formality, so the claim should be explicitly scoped to GPT-4 or discussed as a known trade-off.
  4. [Section 4.2] The alignment vector valign is defined as a difference of centroid differences in embedding space, but the paper does not establish that this vector isolates style rather than content or topic. Because the centroids are computed only on the top/bottom 20% style subsets, the difference µ(L2,S)-µ(L1,S) will also reflect corpus-specific vocabulary and domain drift, not only style. Moreover, the statement in Section 4.3 that the transformed embedding lies 'exactly where it should be... in order for Equation 1 to hold' is a conceptual leap: Eq. (1) concerns quantifier scores, not embedding positions. The method may still work, but the paper needs a more direct validation of the retrieved exemplars' style appropriateness, for example an ablation with random exemplars matched for content similarity.
minor comments (5)
  1. [Abstract / Section 5.1] The abstract and Section 5.1 state that RASTA improves style alignment 'without degrading translation quality,' but the Llama-3.2 results in Table A6 show substantial degradation (e.g., GEMBA -27.2% for intimacy); please scope this claim to GPT-4.
  2. [Figure A2 caption] The caption is copied from Figure A1 and says 'Results show style alignment is worst in non-Western languages' even for the RASTA heatmaps; update the caption to describe the RASTA improvement.
  3. [Section 4.2] The notation uses both S and s for the style variable (e.g., vtrans(L1, L2, s)); please unify to avoid confusion.
  4. [Section 3.2] The definition of 'neutral' as a label in 0.4-0.6 is not justified; please state how this interval was chosen and whether it corresponds to any property of the quantifier calibration.
  5. [Table A5] The per-language 'percent RASTA favored' values are reported without the corresponding sample size or the exact number of majority-vote samples per language; please clarify the denominator.

Circularity Check

2 steps flagged · score 6.0 of 10

RASTA's headline A gains are partly an artifact of evaluating with the same style-quantifier family that its prompt instructs the model to match; the human study is too small and low-agreement to fully break the loop.

  1. self definitional [Section 2, Eq. (2); Section 4.3 and Appendix D, Figure A7]
    "A(L1, L2) = r (C1 (XL1) , C2 (T (XL1))) (2) ... For all of our datasets ... we fine-tune a set of Mistral-7B models ... We train a separate model per language ... This text has a <Style> level of {} out of 1 in {} ... Specifically, the translation should also have a <Style> level of {} out of 1 in {}."

    The alignment metric A is defined as the correlation between the source-style quantifier C1 and the target-style quantifier C2, trained with the same recipe on the same native-speaker datasets. RASTA's prompt then hands the LLM the C1 score of the source text and requires the translation to match that score. An instruction-following model therefore mechanically raises the correlation that defines A, independent of whether human listeners perceive better style alignment. The reported gains (up to 56%) thus partly measure self-consistency of the quantifier family rather than perceived alignment; the independent Prolific study is too weak (3 annotators per language, 30 samples, pairwise agreement 0.167-0.247) to certify the automated gains.

  2. fitted input called prediction [Section 5.1, Table 3 and surrounding text; Appendix F]
    "The solid increase in performance from Baseline 2 also indicates that RASTA better aligns style than what can be done by using a well-written prompt. ... Annotators select RASTA a majority of the time – 61% of the time for politeness, and 63% of the time for formality."

    The claim that RASTA outperforms the preserve-style prompt is evaluated with the same quantifier-based A. RASTA's incremental component over Baseline 2 is the explicit numeric style level taken from C1 and exemplars chosen to match that style level; since A is computed from C1 and C2, supplying the C1 score makes the A increase largely expected. The human preference study is genuinely external, but with only 61-63% preference and pairwise agreement of 0.167-0.247, it is not sufficient to show that the large Table 3 A gains correspond to perceived alignment rather than to alignment with the fitted quantifiers.

full rationale

The paper is not wholly circular: RASTA's embeddings are learned from native corpora, traditional translation metrics (GEMBA, COMET KIWI) are external, and the Prolific preference study is an independent check. However, the central quantitative claim—that RASTA improves style alignment by up to 56%—is measured with a metric that is defined as the correlation between two style quantifiers, while RASTA's own prompt explicitly feeds the source quantifier's score to the model and asks the translation to match it. Since the target quantifier is trained with the same recipe and data, the automated A gain is partly a construction of the evaluation loop rather than evidence about human-perceived style. The paper's Section 8 concession that quantifier RMSE is 0.195 and that style is subjective, together with the low annotator agreement in Appendix F, supports this concern. The Anglocentric-bias and neutrality-bias characterizations also rely on the same quantifiers but are less affected because they are comparative distributional claims. Overall, the derivation chain partially reduces to its own fitted inputs, so the score is 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

RASTA's alignment mapping is a derived embedding vector, not an independent postulated entity; no new particles, forces, or mechanisms are introduced. The main load-bearing additions are the trained style quantifiers and the embedding-space arithmetic assumptions, both listed above.

free parameters (3)
  • 14 per-language per-style style quantifiers (Mistral-7B QLoRA regressors) = trained; test RMSE 0.157-0.289 per style-language
    A, the outcome metric, is defined as correlation between these models' outputs (Eq. 2); RASTA also injects their predicted style label into the prompt, so the quantifiers are load-bearing.
  • number of few-shot exemplars k = 5
    RASTA always retrieves five exemplars; no ablation or tuning curve is reported, and Appendix A6 shows Llama over-relies on them.
  • style-level discretization for centroids (top and bottom 20%) = 0.2 and 0.2 quantiles
    Embedding distances in Table A1 use top and bottom 20% of style scores; the mapping uses style buckets, so the cutoffs influence the alignment vectors.
assumptions (4)
  • domain assumption Per-language style labels are comparable on a single 0-1 scale across languages.
    Eq. (1) requires C1(x)=C2(T(x)); C1 and C2 are trained independently on different languages with no cross-lingual calibration or shared anchors.
  • ad hoc to paper Difference of embedding centroids is a meaningful semantic direction for cultural style shift.
    RASTA defines vnative, vtrans, and valign as vector differences in BGE-M3 space and treats their arithmetic as a cultural correction (Section 4.2); no evidence is given that these directions are linear or linguistically interpretable.
  • domain assumption Style is a static scalar property of a text.
    The metric and RASTA treat style as a single 0-1 label, abstracting away within-culture subjectivity, as acknowledged in Section 8.
  • domain assumption Retrieved exemplars influence the target LLM to preserve style without degrading content.
    RASTA's few-shot mechanism; Appendix A6 shows this fails for Llama-3.2-11B, with large drops in GEMBA and COMET KIWI.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Style Alignment in Cross-Cultural Translation." pith.science (2026). https://pith.science/paper/G3VNJMAL

@misc{pith2026250700216,
  author       = {Pith},
  title        = {Pith review of: Towards Style Alignment in Cross-Cultural Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G3VNJMAL}},
  note         = {Machine review of arXiv:2507.00216}
}
read the original abstract

Successful communication depends on the speaker's intended style (i.e., what the speaker is trying to convey) aligning with the listener's interpreted style (i.e., what the listener perceives). However, cultural differences often lead to misalignment between the two; for example, politeness is often lost in translation. We characterize the ways that LLMs fail to translate style - biasing translations towards neutrality and performing worse in non-Western languages. We mitigate these failures with RASTA (Retrieval-Augmented STylistic Alignment), a method that leverages learned stylistic concepts to encourage LLM translation to appropriately convey cultural communication norms and align style.

Figures

Figures reproduced from arXiv: 2507.00216 by the authors.

Figure 1
Figure 1. An example of cross-cultural communication [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Evaluating how translation affects style. We first select a multilingual style corpus (e.g. the Holistic [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. How well do LLMs align style during translation? We show [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: What languages cause Google Translate to struggle the most when aligning style? We show [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Why do LLMs fail to align style during translation? We plot the politeness distributions of text generated [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: RASTA: our method to align style, shown using English and Japanese. In (1) and (2), we discover [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Comparing style alignment between vanilla [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 36 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]

    Sweta Agrawal, Chunting Zhou, Mike Lewis, Luke Zettlemoyer, and Marjan Ghazvininejad. 2023. https://doi.org/10.18653/v1/2023.findings-acl.564 In-context examples selection for machine translation . In Findings of the Association for Computational Linguistics: ACL 2023, pages 8857--8873, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-acl.564 2023

  4. [4]

    Seyma Albarino. 2024. https://slator.com/translation-companies-accelerating-adoption-large-language-models-alc-survey/ Translation companies accelerating adoption of large language models, alc survey

    work page 2024

  5. [5]

    Douglas Biber and Susan Conrad. 2009. Register, genre, and style. Cambridge University Press

    work page 2009

  6. [6]

    Eleftheria Briakou, Di Lu, Ke Zhang, and Joel Tetreault. 2021. https://doi.org/10.18653/v1/2021.naacl-main.256 Ol \'a , bonjour, salve! XFORMAL : A benchmark for multilingual formality style transfer . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 31...

    doi:10.18653/v1/2021.naacl-main.256 2021

  7. [7]

    Bram Bulte and Arda Tezcan. 2019. Neural fuzzy repair: Integrating fuzzy matches into neural machine translation. In 57th Annual Meeting of the Association-for-Computational-Linguistics (ACL), pages 1800--1809

    work page 2019

  8. [8]

    Yong Cao, Yova Kementchedjhieva, Ruixiang Cui, Antonia Karamolegkou, Li Zhou, Megan Dare, Lucia Donatelli, and Daniel Hershcovich. 2024. Cultural adaptation of recipes. Transactions of the Association for Computational Linguistics, 12:80--99

    work page 2024

Show all 60 references
  1. [9]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024 a . https://arxiv.org/abs/2402.03216 Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . Preprint, arXiv:2402.03216

    2024 arXiv

  2. [10]

    Meiqi Chen, Fandong Meng, Yingxue Zhang, Yan Zhang, and Jie Zhou. 2024 b . Crat: A multi-agent framework for causality-enhanced reflective and retrieval-augmented translation with large language models. arXiv preprint arXiv:2410.21067

    2024 arXiv

  3. [11]

    Simone Conia, Daniel Lee, Min Li, Umar Farooq Minhas, Saloni Potdar, and Yunyao Li. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.914 Towards cross-cultural machine translation with retrieval-augmented generation from multilingual knowledge graphs . In Proceedings of the 2...

    2024 doi

  4. [12]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \' a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1911.02116 Unsupervised cross-lingual representation learning at scale . C...

    2019 arXiv

  5. [13]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36

    2024

  6. [14]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  7. [15]

    Fatima-zahra El-Alami, Said Ouatik El Alaoui, and Noureddine En Nahnahi. 2022. A multilingual offensive language detection method based on transfer learning from transformer fine-tuning model. Journal of King Saud University-Computer and Information Sciences, 34(8):6048--6056

    2022

  8. [16]

    Xavier Garcia, Yamini Bansal, Colin Cherry, George Foster, Maxim Krikun, Melvin Johnson, and Orhan Firat. 2023. The unreasonable effectiveness of few-shot learning for machine translation. In International Conference on Machine Learning, pages 10867--10878. PMLR

    2023

  9. [17]

    Xavier Garcia, Noah Constant, Mandy Guo, and Orhan Firat. 2021. Towards universality in multilingual text rewriting. arXiv preprint arXiv:2107.14749

    2021 arXiv

  10. [18]

    Gemma-Team. 2024. https://arxiv.org/abs/2403.08295 Gemma: Open models based on gemini research and technology . Preprint, arXiv:2403.08295

    2024 arXiv

  11. [19]

    Jiatao Gu, Yong Wang, Kyunghyun Cho, and Victor OK Li. 2018. Search engine guided neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32

    2018

  12. [20]

    Edward T Hall. 1976. Beyond culture. Anchor

    1976

  13. [21]

    Shreya Havaldar, Matthew Pressimone, Eric Wong, and Lyle Ungar. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.419 Comparing styles across languages . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6775--6791, Singapore. Ass...

    2023 doi

  14. [22]

    Zhiwei He, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi, and Xing Wang. 2024. https://doi.org/10.1162/tacl_a_00642 Exploring human-like translation strategy with large language models . Transactions of the Association for Computati...

    2024 doi

  15. [23]

    Daniel Hershcovich, Stella Frank, Heather Lent, Miryam de Lhoneux, Mostafa Abdou, Stephanie Brandl, Emanuele Bugliarello, Laura Cabello Piqueras, Ilias Chalkidis, Ruixiang Cui, Constanza Fierro, Katerina Margatina, Phillip Rust, and Anders S gaard. 2022. https://doi.org/10.186...

    2022 doi

  16. [24]

    Geert Hofstede. 1986. Cultural differences in teaching and learning. International Journal of intercultural relations, 10(3):301--320

    1986

  17. [25]

    Jing Huang and Diyi Yang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.509 Culturally aware natural language inference . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7591--7609, Singapore. Association for Computational Linguistics

    2023 doi

  18. [26]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  19. [27]

    O hman, Piao Hui, and J \

    Kaisla Kajava, Emily \"O hman, Piao Hui, and J \"o rg Tiedemann. 2020. Emotion preservation in translation: Evaluating datasets for annotation projection. In Digital Humanities in the Nordic Countries, pages 38--50. CEUR

    2020

  20. [28]

    Mohamad Ahmad Saleem Khasawneh. 2023. The potential of ai in facilitating cross-cultural communication through translation. Journal of Namibian Studies: History Politics Culture, 37:107--130

    2023

  21. [29]

    Tom Kocmi and Christian Federmann. 2023. https://aclanthology.org/2023.eamt-1.19 Large language models are state-of-the-art evaluators of translation quality . In Proceedings of the 24th Annual Conference of the European Association for Machine Translation, pages 193--203, Tam...

    2023

  22. [30]

    Kalpesh Krishna, Deepak Nathani, Xavier Garcia, Bidisha Samanta, and Partha Talukdar. 2022. https://arxiv.org/abs/2110.07385 Few-shot controllable style transfer for low-resource multilingual settings . Preprint, arXiv:2110.07385

    2022 arXiv

  23. [31]

    Sangmin-Michelle Lee. 2023. The effectiveness of machine translation in foreign language education: a systematic review and meta-analysis. Computer Assisted Language Learning, 36(1-2):103--125

    2023

  24. [32]

    Darrin R Lehman, Chi-yue Chiu, and Mark Schaller. 2004. Psychology and culture. Annu. Rev. Psychol., 55:689--714

    2004

  25. [33]

    Cheng Li, Damien Teney, Linyi Yang, Qingsong Wen, Xing Xie, and Jindong Wang. 2024 a . https://arxiv.org/abs/2405.15145 Culturepark: Boosting cross-cultural understanding in large language models . Preprint, arXiv:2405.15145

    2024 arXiv

  26. [34]

    Shuang Li, Jiangjie Chen, Siyu Yuan, Xinyi Wu, Hao Yang, Shimin Tao, and Yanghua Xiao. 2024 b . Translate meanings, not just words: Idiomkb’s role in optimizing idiomatic translation with language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume...

    2024

  27. [35]

    Angela Moorjani and Thomas T Field. 1988. Semiotic and sociolinguistic paths to understanding culture. In Toward a new integration of language and culture. Reports of the Northeast Conference on the Teaching of Foreign Languages, pages 25--45

    1988

  28. [36]

    Ojha, Akanksha Bansal, Deepak Alok, John P

    Sourabrata Mukherjee, Atul Kr. Ojha, Akanksha Bansal, Deepak Alok, John P. McCrae, and Ondřej Dušek. 2024. https://arxiv.org/abs/2405.20805 Multilingual text style transfer: Datasets & models for indian languages . Preprint, arXiv:2405.20805

    2024 arXiv

  29. [37]

    Maria Nadejde, Anna Currey, Benjamin Hsu, Xing Niu, Marcello Federico, and Georgiana Dinu. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.47 C o C o A - MT : A dataset and benchmark for contrastive controlled MT with application to formality . In Findings of the Associa...

    2022 doi

  30. [38]

    Xing Niu, Marianna Martindale, and Marine Carpuat. 2017. https://doi.org/10.18653/v1/D17-1299 A study of style in machine translation: Controlling the formality of machine translation output . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Proce...

    2017 doi

  31. [39]

    Xing Niu, Sudha Rao, and Marine Carpuat. 2018. Multi-task neural models for translating between styles within and across languages. arXiv preprint arXiv:1806.04357

    2018 arXiv

  32. [40]

    Jiaxin Pei, V \' tor Silva, Maarten Bos, Yozen Liu, Leonardo Neves, David Jurgens, and Francesco Barbieri. 2023. Semeval-2023 task 9: Multilingual tweet intimacy analysis. In Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2235--2246

    2023

  33. [41]

    Denis Peskov, Viktor Hangya, Jordan Boyd-Graber, and Alexander Fraser. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.315 Adapting entities across languages and cultures . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3725--3750, Punta ...

    2021 doi

  34. [42]

    Flor Miriam Plaza-del Arco, Carlo Strapparava, L Alfonso Urena Lopez, and M Teresa Mart \' n-Valdivia. 2020. Emoevent: A multilingual emotion corpus based on different events. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 1492--1498

    2020

  35. [43]

    Sudha Rao and Joel Tetreault. 2018. https://doi.org/10.18653/v1/N18-1012 Dear sir or madam, may I introduce the GYAFC dataset: Corpus, benchmarks and metrics for formality style transfer . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association ...

    2018 doi

  36. [44]

    Ricardo Rei, Marcos Treviso, Nuno M Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, Jos \'e GC De Souza, Taisiya Glushkova, Duarte M Alves, Alon Lavie, et al. 2022. Cometkiwi: Ist-unbabel 2022 submission for the quality estimation shared task. arXiv preprint arXiv...

    2022 arXiv

  37. [45]

    Parker Riley, Noah Constant, Mandy Guo, Girish Kumar, David C Uthus, and Zarana Parekh. 2021. Textsettr: Few-shot text style extraction and tunable targeted restyling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Inter...

    2021

  38. [46]

    Parker Riley, Timothy Dozat, Jan A Botha, Xavier Garcia, Dan Garrette, Jason Riesa, Orhan Firat, and Noah Constant. 2023. Frmt: A benchmark for few-shot region-aware machine translation. Transactions of the Association for Computational Linguistics, 11:671--685

    2023

  39. [47]

    Elijah Rippeth, Sweta Agrawal, and Marine Carpuat. 2022. Controlling translation formality using pre-trained multilingual language models. arXiv preprint arXiv:2205.06644

    2022 arXiv

  40. [48]

    Andrea Schioppa, David Vilar, Artem Sokolov, and Katja Filippova. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.535 Controlling machine translation for multiple attributes with additive interventions . In Proceedings of the 2021 Conference on Empirical Methods in Natural L...

    2021 doi

  41. [49]

    Barbara C Schouten and Ludwien Meeuwesen. 2006. Cultural differences in medical communication: a review of the literature. Patient education and counseling, 64(1-3):21--34

    2006

  42. [50]

    Rico Sennrich, Barry Haddow, and Alexandra Birch-Mayne. 2016. Controlling politeness in neural machine translation via side constraints. In 15th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages...

    2016

  43. [51]

    Rustam Shadiev and Yueh-Min Huang. 2016. Facilitating cross-cultural understanding with learning activities supported by speech-to-text recognition and computer-aided translation. Computers & Education, 98:130--141

    2016

  44. [52]

    Vered Shwartz. 2022. https://doi.org/10.18653/v1/2022.findings-acl.224 Good night at 4 pm?! time expressions in different cultures . In Findings of the Association for Computational Linguistics: ACL 2022, pages 2842--2853, Dublin, Ireland. Association for Computational Linguistics

    2022 doi

  45. [53]

    Anirudh Srinivasan and Eunsol Choi. 2022. Tydip: A dataset for politeness classification in nine typologically diverse languages. arXiv preprint arXiv:2211.16496

    2022 arXiv

  46. [54]

    Deborah Tannen. 1983. Cross-Cultural Communication. ERIC

    1983

  47. [55]

    Jenny A. Thomas. 1983. https://api.semanticscholar.org/CorpusID:35667508 Cross-cultural pragmatic failure . Applied Linguistics, 4:91--112

    1983

  48. [56]

    Enrica Troiano, Roman Klinger, and Sebastian Pad \'o . 2020. Lost in back-translation: Emotion preservation in neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 4340--4354

    2020

  49. [57]

    Jiaan Wang, Fandong Meng, Yingxue Zhang, and Jie Zhou. 2024. Retrieval-augmented machine translation with unstructured knowledge. arXiv preprint arXiv:2412.04342

    2024 arXiv

  50. [58]

    Jitao Xu, Josep-Maria Crego, and Jean Senellart. 2020. Boosting neural machine translation with similar translations. In Annual Meeting of the Association for Computational Linguistics, pages 1570--1579. Association for Computational Linguistics

    2020

  51. [59]

    Binwei Yao, Ming Jiang, Tara Bobinac, Diyi Yang, and Junjie Hu. 2024. https://arxiv.org/abs/2305.14328 Benchmarking machine translation with cultural awareness . Preprint, arXiv:2305.14328

    2024 arXiv

  52. [60]

    Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.176 Multilingual machine translation with large language models: Empirical results and analysis . In Findings of the ...

    2024 doi

Pith tools

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