Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

XCR-Bench: Benchmarking Cross-Cultural Reasoning in LLMs via Culture-Specific Items and Hall's Triad

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper argues that LLMs' cross-cultural competence drops as culture becomes less visible, and that models harbor regional and ethno-religious biases even within a single language.

desk verdict A genuinely useful and honest benchmark resource, but the abstract's headline significance claim about Hall-level declines is not supported by any test in the body, and the trend is metric-dependent. read the letter →

arxiv 2601.14063 v2 pith:W3QA5B4O submitted 2026-01-20 cs.CL cs.AIcs.CY

Mohsinul Kabir , Tasnim Ahmed , Md Mezbaur Rahman , Shaoxiong Ji , Hassan Alhuzali , Yuechen Jiang , Jimin Huang , Sophia Ananiadou This is my paper · ORCID
classification cs.CLcs.AIcs.CY
keywords cross-culturalreasoningculture-specificitemsculturaladaptationvisibilitylevelsLLMevaluationbiasBengaliregionalvariationethno-religious
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

XCR-Bench asks whether LLMs can do more than recall Western cultural facts: can they spot which words in a sentence are culture-specific, predict them in context, and adapt them so the sentence lands naturally in another culture? To test this, the authors built a corpus of 4,136 parallel sentences around 1,098 Western culture-specific items, annotated with category labels and a visibility level (visible practices versus semi-visible and invisible norms, values, and beliefs), plus culturally adapted versions in Chinese, Arabic, and two Bengali variants. Evaluating eight multilingual LLMs, the paper finds that models predict masked culture-specific terms fairly well but identify and adapt them poorly, and that performance declines systematically from visible to invisible cultural content. It also reports that, within Bengali, models adapt more successfully to West Bengal/Hindu-associated forms than to Bangladesh/Muslim-associated forms, evidence of regional and ethno-religious bias inside a single language. If this holds, culture benchmarks that only measure recall or translation overstate what models know, and progress on cultural alignment should be measured at the less visible layers.

What carries the argument

The load-bearing object is the XCR-Bench corpus itself: 4,136 parallel sentences built from 1,098 Western culture-specific items, each annotated with a CSI category and a cultural visibility level from the iceberg/triad model (visible, semi-visible, invisible), and paired with adjudicated cultural adaptations into Chinese, Arabic, and two Bengali variants. Three tasks use it: CSI identification, CSI prediction, and CSI adaptation. The adaptation task forces models to choose an explicit adaptation strategy (e.g., transference, cultural equivalent, neutralization) and to produce both an intra-lingual English adaptation and an inter-lingual target-language adaptation. This combination is what l

What would settle it

Re-annotate a random subset of Bengali adaptation instances with independent expert annotators from Bangladesh and from West Bengal, blind to the current gold, and re-run the eight models; if the West Bengal advantage does not replicate—or flips when judged against Bangladeshi-expert gold—the regional-bias claim is an artifact of the gold, not a property of the models.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-cultural competence in LLMs should be measured as reasoning over culture-specific items, and that when measured this way, contemporary models fail in a structured pattern: they can predict a masked Western cultural term but cannot reliably point to which words are culture-specific; their scores decline as cultural content moves from visible practices to semi-visible and invisible norms and values; and in adapting Western items to Bengali they systematically favor West Bengal and Hindu-associated forms (e.g., festival terms, kinship terms) over Bangladesh and Muslim-associated ones, even though Bengali Muslims form a majority of speakers. The reported d

Load-bearing premise

The load-bearing premise is that the gold annotations—especially the cross-cultural adaptation pairs and the cultural-visibility labels, which reach only 0.64 inter-annotator agreement—are culturally valid ground truth; if those choices are systematically skewed, the model rankings and the regional-bias conclusion would be artifacts of annotation rather than properties of the models.

Editorial extensions

If this is right

  • Cultural evaluation must include identification and adaptation tasks, because prediction alone overstates competence.
  • Benchmarks should stratify by cultural visibility level; a model that handles visible practices well can still fail on norms, values, and beliefs.
  • Adaptation quality is not translation quality: models do better when producing the target language than when adapting within English, and do better on cultural stereotypes than on grounded cultural references.
  • Even a single language is not one culture; region-level splits such as Bengali expose biases that language-level benchmarks hide.
  • Weighting benchmark scores by cultural visibility level would yield different model rankings than averaging across all items.

Reading between the lines

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

  • A testable extension: apply the same three tasks to other pluricentric languages (e.g., Spanish, Portuguese, Arabic); if the visible-to-invisible decline and regional bias recur, they reflect general properties of how LLMs store cultural knowledge rather than quirks of one corpus.
  • The 'Non-transferable' adaptation category could be reused as a refusal-safety test for culturally sensitive content, checking whether a model knows when adaptation would violate local norms instead of forcing an equivalent.
  • The intra- versus inter-lingual gap suggests a practical deployment lesson worth testing directly: asking a model to translate into the target language may preserve culture better than asking it to 'localize' in English.
  • Because the Bengali bias emerges despite Bangladesh being the larger speaker population, sampling-based evaluations of cultural competence should weight sub-communities by demographics and dialect, not just by language name.
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

3 major / 5 minor

Summary. XCR-Bench introduces a human-annotated benchmark for evaluating cross-cultural reasoning in LLMs. The corpus contains 4,136 parallel sentences with 1,098 annotated Culture-Specific Items (CSIs), mapped to Newmark's CSI taxonomy and Hall's Triad of Culture (Visible/Semi-visible/Invisible). Three tasks are defined: CSI Identification, CSI Prediction, and CSI Adaptation, with adaptation pairs for Chinese, Arabic, and two Bengali regional variants (West Bengal and Bangladesh). Eight multilingual LLMs are evaluated. The paper reports that models are weak at identifying CSIs, particularly for Social Etiquette and Cultural Reference, that performance declines systematically from Visible to Invisible cultural levels, and that Bengali adaptation shows a West Bengal/Hindu-associated bias over Bangladesh/Muslim-associated forms.

Significance. If the resource and findings hold, XCR-Bench is a useful contribution: it is the first benchmark to combine Newmark's CSI categories with Hall's triadic levels, it covers three reasoning tasks rather than translation-only settings, and it explicitly separates Bengali regional variants, enabling fine-grained bias analysis. Strengths include the public release of corpus and code, human annotation with documented qualifications and adjudication, and the use of both hard and soft metrics. The central empirical claims, however, are currently stronger than the evidence in the body of the paper, and the Bengali bias conclusion rests on small effect sizes. The dataset itself is plausible and valuable, but the advertised headline results need to be either properly supported or appropriately qualified.

major comments (3)
  1. [Abstract and §4.1] The abstract states 'Performance declines significantly ... deeper cultural levels (p<0.005, 8/8 models)', but no statistical test, test statistic, or p-value for this claim appears anywhere in §4.1, Appendix G, or the metric appendix. §4.1 only cites Figures 5 and 6 and Table 9. The figures do not uniformly support a monotonic decline: in Fig. 5a, HI-CSI is lowest at Semi-visible rather than Invisible, and in Fig. 6a HP-CSI is described as 'mixed'. Since the Hall-level annotations have kappa 0.64 (§2.1), the claim as stated is unsupported. Please either provide a defined significance test (e.g., a mixed-effects model or paired test across models and items), report per-metric results, and address label noise, or remove the 'p<0.005, 8/8 models' assertion from the abstract.
  2. [Table 4 and §4.2] The regional-bias conclusion in §4.2 is based on mean differences of 0.001–0.030 in CSIbert/SENTbert units. Sixteen one-sided Wilcoxon tests are performed (8 models × intra/inter), of which only four reach p<0.05 or p<0.01, with no multiple-comparison correction. The qualitative examples (Puja vs Eid, dada vs bhai) are suggestive, but the claim of 'pronounced regional and ethno-religious biases' is stronger than the quantitative evidence. Please report effect sizes, correct for multiple comparisons, and quantify how often models choose West Bengal- vs Bangladesh-associated forms on the actual adaptation items.
  3. [§2.1 and §4.1] The base corpus sentences are generated by GPT-4o, Claude-3.7-Sonnet, and DeepSeek-R1, with human selection among three candidates. The same three models are then evaluated on these sentences in all three tasks. This creates a potential contamination or self-generation advantage: a model may perform better on sentences it generated (and whose CSI it saw at generation time) than on sentences from other sources. The paper does not report the distribution of selected sentences by generator or analyze whether model rankings change when restricting to sentences not generated by that model. This is load-bearing for the comparative model evaluation and should be addressed, at least as a robustness check.
minor comments (5)
  1. [Abstract vs §2.3] The abstract in the full text says '4.9k parallel sentences' while the supplied arXiv abstract says '4.1k' and §2.3 says '4,136'. Please harmonize these numbers.
  2. [§2.1] The kappa values (0.68 for CSI categories, 0.64 for Hall levels) are reported without confidence intervals or per-category breakdowns. Given the central role of Hall-level annotation, a brief discussion of disagreement patterns would be helpful.
  3. [Table 9] Several Hall-level cultural elements, e.g., 'Orientations', contain very few instances and many 0% scores across models. Reporting scores for such small cells without instance counts is misleading; please add counts or suppress low-support rows.
  4. [General presentation] The paper contains numerous OCR-like artifacts ('7aragraph', '⚶', 'oculturep', 'Cohenns', 'uni00AD'), which should be cleaned. Some tables and figures are hard to read in the provided text, e.g., Figure 4's heatmap labels are cut off.
  5. [Appendix F.3] The inter-lingual BERTScore evaluation uses language-specific BERT models ('bert-base-arabic', 'chinese-bert-wwm-ext', 'bangla-bert-base') for the target side and a multilingual model for cross-language pairs. This is reasonable, but the choice of model for intra-lingual English (bert-base-uncased) may not be ideal for culturally adapted English; please justify or note the limitation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark answers come from external resources and human adjudication, not from the evaluated models.

full rationale

XCR-Bench's gold standard is externally grounded: CSIs are extracted from CANDLE and Cultural Atlas, contextualized via Wikipedia, and the CSI categories, Hall levels, and cross-cultural adaptation pairs are produced by human annotation and adjudication. The evaluated LLMs are never used to create the gold answers; LLM-generated candidate sentences are filtered and selected by human annotators, so no model output feeds back into the target labels. The three tasks (CSI Identification, CSI Prediction, CSI Adaptation) score model outputs against these fixed human gold labels, and no parameter is fitted to the evaluation data. The paper's self-citations (Kabir et al. 2025a,b,c) appear only in background discussion and are accompanied by external citations (e.g., Navigli et al. 2023, Röttger et al. 2024); they are not load-bearing for the benchmark's validity and no uniqueness claim is imported from the authors' own prior work. The abstract's 'p<0.005, 8/8 models' claim is not supported by any reported statistical test in the body, and the Hall-level trend is metric-dependent (Fig. 5a shows HI-CSI lowest at Semi-visible), but this is an evidentiary/reporting gap rather than a circular derivation. The explicit Limitations section (limited Newmark categories, limited cultures, single prompting strategy) also does not indicate circularity. Overall, no step in the paper reduces by construction to its own inputs, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new theoretical entities (no new forces/particles); all assumptions are domain assumptions about cultural taxonomies and evaluation metrics. The main load is carried by the validity of Newmark and Hall frameworks and the annotation reliability, with Hall-level agreement below typical high-stakes thresholds.

assumptions (7)
  • domain assumption Newmark's five CSI categories form a valid typology for labeling culture-specific items.
    The entire annotation scheme relies on Newmark (1988) as a ground-truth taxonomy; no independent justification for this specific categorization is given in the paper.
  • domain assumption Hall's Triad (Visible/Semi-visible/Invisible) is a valid decomposition of cultural knowledge into three levels that can be assigned reliably to CSIs.
    Used throughout for the 'cultural depth' analysis; inter-annotator agreement is only kappa=0.64 for this assignment, indicating substantial but not high reliability.
  • domain assumption The mapping between Newmark categories, Liu et al. taxonomy, and Hall's levels (Tables 5 and 8) is correct and contextually appropriate.
    The paper acknowledges a CSI's category can shift with context, so the mapping is interpretative; errors here would propagate into all category-level claims.
  • domain assumption LLM-generated sentences (by GPT-4o, Claude-3.7, DeepSeek-R1) that pass human selection are realistic enough to serve as evaluation stimuli for everyday communication.
    Step 2 of corpus construction relies on LLM generation; these sentences may carry generation artifacts that favor certain models in subsequent evaluation.
  • domain assumption The four adaptation equivalence types (direct, functional, neutral, non-transferable) are sufficient to represent all valid cultural adaptations.
    Used as the annotation guideline for parallel corpus construction; the paper treats cultural pluralism as acceptable, which makes the gold set one of several possible valid sets.
  • domain assumption BERTScore with language-specific BERT models gives a meaningful measure of semantic similarity for CSI adaptation quality.
    The evaluation metrics in Section 3.4 and Appendix F rely on BERTScore; this is a standard but imperfect proxy, especially for low-resource Bengali variants and cross-lingual comparisons.
  • standard math A one-sided Wilcoxon signed-rank test on per-CSI-category performance differences is a valid way to test the regional bias claim.
    Table 4 uses this test; it assumes paired data and is appropriate for ordinal scores, but the paper does not state whether multiple-comparison corrections were applied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XCR-Bench: Benchmarking Cross-Cultural Reasoning in LLMs via Culture-Specific Items and Hall's Triad." pith.science (2026). https://pith.science/paper/W3QA5B4O

@misc{pith2026260114063,
  author       = {Pith},
  title        = {Pith review of: XCR-Bench: Benchmarking Cross-Cultural Reasoning in LLMs via Culture-Specific Items and Hall's Triad},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3QA5B4O}},
  note         = {Machine review of arXiv:2601.14063}
}
read the original abstract

Cross-cultural competence in large language models (LLMs) requires understanding and adapting Culture-Specific Items (CSIs) across varying cultural contexts. However, progress in evaluating this capability remains limited by the lack of high-quality CSI-annotated corpora with parallel cross-cultural sentence pairs. We introduce XCR-Bench, a Cross(X)-Cultural Reasoning Benchmark containing 4.1k parallel sentences and 1,098 CSIs across three reasoning tasks. XCR-Bench integrates Newmark's CSI framework with Hall's Triad of Culture, enabling evaluation across levels of cultural visibility -- from observable practices to implicit social norms and values. Experiments on eight multilingual LLMs show that state-of-the-art models exhibit consistent weaknesses in identifying and adapting specific categories of CSIs, revealing a gap between surface-level recall and explicit cultural reasoning. Performance declines significantly on culturally sensitive categories and deeper cultural levels (p<0.005, 8/8 models), and adaptation quality varies systematically across target cultures and Bengali regional variants, indicating encoded regional and ethno-religious biases even within a single linguistic setting. We publicly release the corpus and code to support future research on cross-cultural NLP.

Figures

Figures reproduced from arXiv: 2601.14063 by the authors.

Figure 1
Figure 1. XCR-Bench corpus creation and annotation pipeline. (1) Culture-specific items (CSIs) and their contexts are extracted from well-known cultural databases. (2) Sentences are generated using LLMs and filtered for accuracy and fluency. (3) Each sentence is annotated with a CSI category and mapped to Hall’s cultural levels. allel corpus. 2.1 Construction of Base XCR-Bench Corpus [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance of LLMs across CSI categories for CSI Identification and Prediction tasks. For clarity, only [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Mapping between CSI categories and Hall’s cultural levels. The textured bar plots indicate a higher [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Mapping between CSI categories and Hall’s cultural elements in the XCR-Bench Corpus. [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Performance of LLMs across Hall’s Triad of Culture (visible, semi-visible, and invisible) for CSI Iden [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Performance of LLMs across Hall’s Triad of Culture (visible, semi-visible, and invisible) for CSI Predic [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Inter–Intra CSI adaptation gain for Arabic and Chinese. The heatmap reports the difference between inter-lingual and intra-lingual CSI adaptation scores (Inter − Intra) under the CSIbert metric across models and CSI categories. Positive gains across most CSI categories…
Figure 8
Figure 8. Figure 8: Inter–Intra CSI adaptation gain for Bengali (West Bengal and Bangladesh). The heatmap reports the difference between inter-lingual and intra-lingual CSI adaptation scores (Inter − Intra) under the CSIbert metric across models and CSI categories. Positive values indicat…
Figure 9
Figure 9. Figure 9: Performance of LLMs on CSI Adaptation (CSI [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Performance of LLMs on CSI Adaptation (CSI [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]
Figure 11
Figure 11. Figure 11: Evaluation Prompt for CSI Identification Task. [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]
Figure 12
Figure 12. Figure 12: Evaluation Prompt for CSI Prediction Task. [PITH_FULL_IMAGE:figures/full_fig_p029_12.png]
Figure 13
Figure 13. Figure 13: Evaluation Prompt for CSI Adaptation Task. [PITH_FULL_IMAGE:figures/full_fig_p030_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CultureForest: Understanding and Evaluating Cultural Norm Grounded Reasoning in LLMs

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    CultureForest benchmark shows top LLMs degrade sharply on open-ended cultural reasoning tasks, exhibit regional disparities, and are limited more by effective use of knowledge than by lack of knowledge itself.

Reference graph

Works this paper leans on

38 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    The original CSI is retained un- changed in the adapted output (e.g., sari, ki- mono)

    Transference. The original CSI is retained un- changed in the adapted output (e.g., sari, ki- mono)

  2. [2]

    A culturally analogous termfromthetargetcultureissubstitutedtocon- vey a similar function or social meaning (e.g., adapting Thanksgiving as a local harvest festi- val)

    Cultural Equivalent. A culturally analogous termfromthetargetcultureissubstitutedtocon- vey a similar function or social meaning (e.g., adapting Thanksgiving as a local harvest festi- val)

  3. [3]

    a traditional Japanese robe

    Neutralization. The CSI is replaced with a de- scription that explains its function or meaning in culturally neutral terms (e.g., rendering ki- mono as “a traditional Japanese robe”)

  4. [4]

    Fed- eral Parliament

    Literal Translation. The CSI is translated di- rectly on a word-by-word basis into the target language (e.g., translating Bundestag as “Fed- eral Parliament”)

  5. [5]

    In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2054–2059, Barcelona (online).InternationalCommitteeforComputational Linguistics

    KUISAIL at SemEval-2020 task 12: BERT- CNNforoffensivespeechidentificationin socialme- dia. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2054–2059, Barcelona (online).InternationalCommitteeforComputational Linguistics. Sougata Saha, Saurabh Kumar Pandey, Harshit Gupta, andMonojitChoudhury.2025. Readingbetweenthe lines: Can llms ...

    arXiv 2020

  6. [6]

    TheCSIisadaptedtoconform to the spelling or pronunciation conventions of the target language (e.g.,Pharisees)

    Naturalization. TheCSIisadaptedtoconform to the spelling or pronunciation conventions of the target language (e.g.,Pharisees)

  7. [7]

    arXiv preprint arXiv:2406.14504

    Translating across cultures: Llms for intralingual cultural adaptation. arXiv preprint arXiv:2406.14504. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie- Yan Liu. 2020. Mpnet: Masked and permuted pre- training for language understanding. Advances in neural information processing systems , 33:16857– 16867. YuchenSong, AndongChen, WenxinZhu, KehaiChen, X...

    arXiv 2020

  8. [8]

    YuhangWang, YanxuZhu, ChaoKong, Shuyu Wei, Xi- aoyuan Yi, Xing Xie, and Jitao Sang

    Cultural influences on word meanings re- vealed through large-scale semantic alignment.Na- ture Human Behaviour, 4(10):1029–1038. YuhangWang, YanxuZhu, ChaoKong, Shuyu Wei, Xi- aoyuan Yi, Xing Xie, and Jitao Sang. 2024. Cdeval: A benchmark for measuring the cultural dimensions of large language models. InProceedings of the 2nd Workshop on Cross-Cultural C...

    2024

Show all 38 references
  1. [9]

    Two adaptation strategies are com- bined, such as retaining the original CSI while also providing an explanatory description

    Couplet. Two adaptation strategies are com- bined, such as retaining the original CSI while also providing an explanatory description

  2. [10]

    A widely recognized and conventionally accepted trans- lation is used (e.g.,Holy See for Saint-Siège)

    Accepted Standard Translation. A widely recognized and conventionally accepted trans- lation is used (e.g.,Holy See for Saint-Siège)

  3. [11]

    A longer explanatory paraphrase or gloss is added, either inline or as afootnote, toclarifytheCSI’smeaningandcul- tural context

    Paraphrase or Gloss. A longer explanatory paraphrase or gloss is added, either inline or as afootnote, toclarifytheCSI’smeaningandcul- tural context

  4. [12]

    the Japanese game of Go

    Classifier. A general category term is added to situate the CSI within a familiar conceptual class (e.g., “the Japanese game of Go”). D Annotation and Annotator Details Annotation Training. The annotation process was preceded by structured training sessions con- ducted by the ...

  5. [14]

    The CSI is retained or adapted and accompanied by a brief explanatory label that clarifies its cultural role or category

    Labeling. The CSI is retained or adapted and accompanied by a brief explanatory label that clarifies its cultural role or category

  6. [16]

    a summer house for wealthy people

    Componential Analysis. The CSI is decom- posed into its constituent semantic components, each of which is explicitly explained (e.g., ren- dering dacha as “a summer house for wealthy people”)

  7. [17]

    The CSI is omitted entirely when it is non-essential to the meaning or when adapta- tion would introduce unnecessary complexity

    Deletion. The CSI is omitted entirely when it is non-essential to the meaning or when adapta- tion would introduce unnecessary complexity

  8. [22]

    • Use natural and idiomatic language in both intra- and inter-lingual adaptations

    Purpose The goal of annotation is to ensure that CSI adap- tations: • Reflect authentic cultural values, norms, and practices of the target culture. • Use natural and idiomatic language in both intra- and inter-lingual adaptations. • Respect cultural, religious, and social sen...

  9. [23]

    [For illustration, adaptation examples are given here in Arabic.] 7aragraphA

    Adaptation Rules Following Newmark’s cultural equivalence frame- work (Newmark, 1988), annotate each CSI using one of the following strategies. [For illustration, adaptation examples are given here in Arabic.] 7aragraphA. Direct Equivalent Rule: Replace theCSIwithaculturallyid...

    1988

  10. [24]

    • Respect religious norms and social conven- tions

    Cultural Sensitivity While doing the annotation, please ensure to: • Avoid culturally inappropriate or taboo con- tent (e.g., alcohol, gambling, or sensitive rela- tionships where applicable). • Respect religious norms and social conven- tions

  11. [25]

    Annotation Checklist For each sentence, please verify: • Cultural accuracy: Is the CSI adapted ap- propriately for the target culture? • Linguistic quality: Is the adaptation natural and idiomatic? • Functional equivalence: Does the adapta- tion preserve the original communica...

  12. [26]

    • Flag creative or uncertain adaptations for fur- ther review

    Final Notes • Consulttheexpertannotatorforambiguousor borderline cases. • Flag creative or uncertain adaptations for fur- ther review. • Use CANDLE and Cultural Atlas as refer- ence sources for relevant cultural practices, rituals, and traditions. 16 Social Tradition Social Et...

    2020

  13. [27]

    Transference: Keep the original word unchanged in the adaptation

  14. [28]

    Cultural Equivalent: Use a similar term from the target culture

  15. [29]

    Neutralization: Explain what the term means or does

  16. [30]

    Literal Translation: Translating the word directly to target culture

  17. [31]

    Label: Add a brief explanation or tag to the term

  18. [32]

    Naturalization: Adapt the word to fit the target language’s spelling or sound

  19. [33]

    Componential Analysis: Break the term into parts and explain each

  20. [34]

    Deletion: Remove unnecessary words or phrases

  21. [35]

    Couplet: Combine two adaptation methods

  22. [36]

    Accepted Standard Translation: Use a commonly accepted adaptation

  23. [37]

    Paraphrase or Gloss: Give a longer explanation or footnote

  24. [38]

    Your task is to adapt the following sentences containing <CSI> tags into Bengali/Chinese/Arabic culture

    Classifier: Add a general category to clarify the termns meaning. Your task is to adapt the following sentences containing <CSI> tags into Bengali/Chinese/Arabic culture. Replace inside the <CSI> tags with culturally relevant practices, behaviors, or terms. Do not remove the <...

  25. [238]

    Pushpdeep Singh, Mayur Patidar, and Lovekesh Vig

    Routledge. Pushpdeep Singh, Mayur Patidar, and Lovekesh Vig

  26. [2020]

    arXiv preprint arXiv:2011.03287

    The apposcorpus: A new multilingual, multi- domain dataset for factual appositive generation. arXiv preprint arXiv:2011.03287. Julia Kharchenko, Tanya Roosta, Aman Chadha, and Chirag Shah. 2024. How well do llms represent values across cultures? empirical analysis of llm respo...

    2011 arXiv

  27. [2022]

    arXiv preprint arXiv:2203.10020

    Challengesandstrategiesincross-culturalnlp. arXiv preprint arXiv:2203.10020. Geert Hofstede, Gert Jan Hofstede, and Michael Minkov. 2010. Cultures et organisations: Nos pro- grammations mentales. Pearson Education France. Jie Huang and Kevin Chen-Chuan Chang. 2023. To- wards r...

    2010 arXiv

  28. [2023]

    arXiv preprint arXiv:2305.14688

    Expertprompting: Instructing large language models to be distinguished experts. arXiv preprint arXiv:2305.14688. Binwei Yao, Ming Jiang, Tara Bobinac, Diyi Yang, and Junjie Hu. 2024. Benchmarking machine translation with cultural awareness. In Findings of the Associ- ation for...

    2024 arXiv

  29. [2024]

    InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 16366–16393

    Having beer after prayer? measuring cultural bias in large language models. InProceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 16366–16393. Roberto Navigli, Simone Conia, and Björn Ross. 2023. Biases in la...

    2023 arXiv

  30. [2025]

    Transac- tions of the Association for Computational Linguis- tics, 13:652–689

    Culturally aware and adapted NLP: A taxon- omy and a survey of the state of the art. Transac- tions of the Association for Computational Linguis- tics, 13:652–689. Tarek Naous, Michael J Ryan, Alan Ritter, and Wei Xu

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.