PaperlayerView exact revision on arXiv

Source-linked AI summary

Supporting and Performing Culture from the Inside

Lea Frermann, Steven Bird

arXiv:2609.32281v1cs.CLcs.AIcs.CY

TL;DR

Cultural research in NLP is shaped by researchers’ cultural perspectives, while textual records may reflect outsiders’ conventions and offer limited access to underlying meanings. The paper applies the emic/etic distinction to survey cultural NLP, finding that most surveyed papers focus on intrinsic cultural knowledge; it argues for more emic approaches.

  • Problem

    Researchers’ cultural perspectives shape cultural NLP, and textual records for many cultures reflect outsiders’ conventions and may not capture underlying meanings reliably.

    Source: §1 · p. 1, §2.2 · p. 3

  • Method

    The paper applies the emic/etic framework to survey cultural NLP and examine assumptions across the Four Stages of AI Development.

    Source: §5 · p. 5

  • Results

    78 papers (73%) focused on assessing or augmenting LLM cultural knowledge, while the survey found no papers co-developing tasks with a culture’s members.

    Source: §2.1 · p. 3

  • Takeaways & Limitations

    The authors suggest moving toward an emic research paradigm, including local language-model deployment using insider sources while recognizing textual-source limitations.

    Source: §3 · p. 4, §5 · p. 5

    The literature survey is representative but not exhaustive, and may have missed a small number of more emic approaches.

    Source: §6 · p. 6

Abstract

from arXiv · show

Anthropology and related disciplines which study culture have often found it useful to consider their epistemological and methodological approaches in terms of emic versus etic. This is the distinction between the insider perspective, how the world is viewed by a member of the culture, versus the outsider perspective, the scientific cataloguing of cultural knowledge and practice. We adopt this framework to systematically analyse assumptions in recent research on culture in NLP along the full pipeline of task selection, data collection, system design and evaluation. We draw attention to the `culture of NLP' which shapes the research focus and approaches of our field, and suggest pathways towards a more culturally attuned AI.

1 Introduction

The paper frames cultural research in NLP through the insider-oriented emic and outsider-oriented etic distinction, arguing that the field’s own culture shapes research and systems. It surveys cultural NLP across the Four Stages of AI Development to identify etic dominance, hidden assumptions, and paths toward an emic agenda.

1 Introduction

  • NLP culture shapes research focus and approaches, including what is observed, how it is interpreted, and what text or speech enters an LLM.It is also embedded in LLM representational choices and chatbot interactional styles, and in evaluations of cultural capability.Source: §1 · p. 1
  • Etic approaches classify observable surface forms from an outsider perspective, while emic approaches focus on the culture-specific underlying system from an insider perspective.The paper relates observable behavior to significance accessible only to insiders.Source: §1 · pp. 1–2
  • The paper aims to challenge seemingly inevitable NLP research and system-development methods and encourage reflection on how NLP culture shapes the field’s goals and contributions.Source: §1 · p. 2
  • The paper surveys cultural NLP across the Four Stages of AI Development, identifies etic dominance and hidden assumptions, and sketches an emic research agenda.It also points to gaps in work that assumes the goal is to develop systems valuable for the culture in question.Source: §1 · p. 2

2 The Etic/Emic Framework

The paper surveys culture-focused NLP through an emic/etic framework and finds etic approaches across AI development stages. It uses the framework to identify assumptions and gaps in task selection, data, system design, and evaluation.

2 The Etic/Emic Framework

  • Etic development treats tasks as general-purpose solutions and text as primary data, while emic approaches center local meanings and culturally meaningful ends.The section contrasts this etic pattern across task definition, data, system design, and deployment and evaluation.Source: §2 · p. 2, §2.4 · p. 4

2.1 Tasks

  • The survey found no papers that co-developed tasks with a cultural community, while 78 papers (73%) focused on cultural knowledge in LLMs.Other surveyed tasks were typically selected and formalized by researchers aiming for generalizability across cultures.Source: §2.1 · p. 3

2.2 Data

  • For many cultures, textual records reflect outsider conventions, and observation, attributed meaning, and textual rendering each mediate what is represented.The paper argues that removing outsider texts does not remove the cultural conditions shaping writing and observation.Source: §2.2 · p. 3
  • 33 papers created cultural benchmarks at least partly synthetically, while 19 drew on community input and English-benchmark translation appeared only three times.The paper notes that LLM-generated concepts or context filtered by humans can distort an emic, intra-cultural perspective.Source: §2.2 · p. 3

2.3 System design

  • Chatbots designed as domain experts risk replacing domain-shaped learning interactions with novice-shaped prompt-response exchanges.The paper presents augmentative agents as an alternative to replicating or replacing human expertise.Source: §2.3 · pp. 3–4
  • 73% of surveyed papers focused on intrinsic evaluation and did not consider a downstream system; task-oriented work followed established NLP designs.The review covers 107 papers published in *ACL or EMNLP from 2022 to April 2026 that mention culture in the title.Source: §2 · p. 2, §2.3 · p. 4

2.4 Deployment and evaluation

  • None of the surveyed task-oriented papers deployed a system in a community for ecologically valid evaluation, and human studies used researcher-defined variables.Those studies were therefore constrained in focus and context and insensitive to user considerations beyond the predefined variables.Source: §2.4 · p. 4

3 Paths Forward

The paper sketches three routes toward emic NLP: local adaptation, community-led local control, and relational approaches that rethink intelligence and AI development.

Local adaptation

  • Local adaptation deploys language models within a single cultural area, using insider sources to make users’ cultural assumptions more likely to align with textual tropes.It remains etic because it relies on textual sources, but may mitigate harms from outsider views and homogenisation toward dominant norms.Source: §3 · p. 4

Local control

  • Local control centers community aspirations, potentially reshaping chatbot interaction around local knowledge transmission, such as teacher-led rather than learner-led initiatives.It can begin with tasks such as text entry and literacy.Source: §3 · p. 4

Relational AI

  • Relational AI rejects one-size-fits-all thinking and essentialised culture, anticipating parochial and hybrid solutions rather than systems based on replicating observable human behaviour.The proposed path develops both technologists’ capacity to engage respectfully with diverse cultures and communities’ capacity to conceive technologies serving local agendas.Source: §3 · pp. 4–5

4 Discussion

The discussion argues for greater emphasis on emic approaches and responds to objections about cultural generalization, technical scope, and identifying insider perspectives.

Counter-argument (a)

  • Intermediate positions between etic and emic imagine culture as learnable idioms or alignable values, while the authors contend that emic development stages are tightly coupled.The discussion frames the distinction as near-categorical rather than a spectrum.Source: §4 · p. 5

Discussion

  • The authors treat etic and emic as complementary standpoints but argue that culture research in NLP would benefit from greater emphasis on emic approaches.They describe much work to date as etic.Source: §4 · p. 5

Counter-argument (b)

  • An emic view considers one culture at a time, while majority-skewed training data and translated or generated additions can make majority culture dominant downstream.The authors say this compromises technology’s attunement to the target culture’s needs and goals.Source: §4 · p. 5

Counter-argument (c)

  • Treating technology as culturally neutral and leaving scenario integration to others conflicts with the view that technological artefacts are social constructs.The discussion points to Science and Technology Studies and work on linguistic and cultural diversity.Source: §4 · p. 5

Counter-argument (d)

  • Author demographics would not establish insider perspective: minority-culture researchers whose work reaches ACL may already have assimilated NLP culture.The authors note that Masakhane and Lanfrica focus on public datasets and reproducing GenAI results from dominant languages.Source: §4 · p. 5

5 Conclusion

The conclusion applies the emic/etic framework to cultural NLP and argues that textual-record limits prevent etic methods from resolving cultural alignment.

Conclusion

  • The review identifies limits intrinsic to etic approaches because layers of culture are absent from or cannot be removed from the textual record accessible to LLMs.The paper applies the framework through a systemic review of cultural NLP literature.Source: §5 · p. 5
  • Manipulating LLM representations or chatbot interactional style does not solve cultural alignment, which the authors describe as society-complete and differing across cultures.The conclusion links the problem’s scope to differences among societies and cultures.Source: §5 · p. 5
  • The authors propose moving toward a more emic paradigm for NLP research.This follows their account of the limits of etic approaches to cultural alignment.Source: §5 · p. 5

6 Limitations

The authors qualify their survey’s scope and acknowledge that their own cultural perspectives shaped the paper.

Limitations

  • The authors’ cultural perspectives shaped the paper.This is an acknowledged feature of the work.Source: §6 · p. 6
  • The literature survey is representative but not exhaustive, and may have missed a small number of more emic approaches.The authors maintain that their main conclusions still hold.Source: §6 · p. 6

A Literature Survey

The survey offers a systematic, non-exhaustive review of recent NLP papers on culture, using manual inspection and labels to characterize their contributions, data creation, and task focus.

A Literature Survey

  • The authors describe the review as systematic but not exhaustive, with broader follow-up survey details in Section A.1.The full reviewed-paper list and labels are provided on GitHub, and summary statistics appear in Tables A.1–A.3.Source: A Literature Survey · p. 9
  • The survey reviewed 107 papers from the main or findings tracks of *ACL and EMNLP, published from 2022 through April 2026, that mentioned culture or cultural in their titles.Multimodal approaches and papers whose primary focus was not culture were excluded.Source: A Literature Survey · p. 9
  • The authors manually labelled papers by primary contribution and, for empirical papers, by data-creation approach and whether the task assessed cultural knowledge in LLMs or proposed culturally sensitive applications.The task distinction is between upstream or intrinsic assessment/augmentation and downstream or extrinsic applications.Source: A Literature Survey · p. 9

A.1 Expanded Survey

The expanded survey searched ACL Anthology papers mentioning culture or cultural and used a keyword search to estimate the presence of emic approaches. Its rough estimate is explicitly qualified by noisy search terms and coverage.

A.1 Expanded Survey

  • The expanded search covered 1,536 ACL Anthology papers mentioning culture or cultural in the title or abstract, then identified 108 mentioning at least one keyword associated with an emic approach.The expanded keyword list included terms such as emic, community-centered, anthropological, insider, and outsider.Source: §A.1 · p. 9
  • If papers matching the expanded keywords take an emic approach, they would make up 7% of the sample; the authors call this a crude approximation because both numerator and denominator are noisy.Source: §A.1 · p. 9
  • Appendix tables report distributions of surveyed papers by main contribution, focus on cultural knowledge in LLMs, and dataset-creation approach.The dataset-creation table excludes papers that did not contribute datasets and those investigating culture in LLM output rather than human artifacts.Source: Table A.1 · p. 9, Table A.2 · p. 9, Table A.3 · p. 9
Loading 2609.32281v1…