PaperlayerView exact revision on arXiv

Source-linked AI summary

Culture is Not Trivia: Sociocultural Theory for Cultural NLP

Naitian Zhou, David Bamman, Isaac L. Bleaman

arXiv:2502.12057v2cs.CLcs.CY

TL;DR

Cultural NLP has advanced without a shared conception of culture, relying instead on varied proxies. The paper applies sociocultural linguistics to clarify what cultural knowledge claims capture and to suggest alternative framings; it concludes that culture should be modeled as dynamic process and that localization may better frame many current projects.

  • Problem

    Cultural NLP relies on varied cultural proxies without a shared conception of culture, while recurring limitations raise unresolved theoretical questions about how to model culture.

    Source: §3 · p. 3

  • Method

    The paper draws on sociocultural linguistics, including an indexicality case study, to clarify cultural knowledge and stereotypes and consider paths forward.

    Source: §1 · p. 2

  • Results

    The paper concludes that metapragmatic discourse primarily reveals stereotypes, and that culturally competent systems require a dynamic model of culture as process rather than trivia.

    Source: §5.1 · p. 6, §7 · p. 9

  • Takeaways & Limitations

    For many current projects, localization may be a more useful and immediate framing than cultural competence or understanding.

    Source: §6.2 · p. 8

    The paper is illustrative rather than an exhaustive survey of cultural NLP.

    Source: §8 · p. 9

Abstract

from arXiv · show

The field of cultural NLP has recently experienced rapid growth, driven by a pressing need to ensure that language technologies are effective and safe across a pluralistic user base. This work has largely progressed without a shared conception of culture, instead choosing to rely on a wide array of cultural proxies. However, this leads to a number of recurring limitations: coarse national boundaries fail to capture nuanced differences that lay within them, limited coverage restricts datasets to only a subset of usually highly-represented cultures, and a lack of dynamicity results in static cultural benchmarks that do not change as culture evolves. In this position paper, we argue that these methodological limitations are symptomatic of a theoretical gap. We draw on a well-developed theory of culture from sociocultural linguistics to fill this gap by 1) demonstrating in a case study how it can clarify methodological constraints and affordances, 2) offering theoretically-motivated paths forward to achieving cultural competence, and 3) arguing that localization is a more useful framing for the goals of much current work in cultural NLP.

1 Introduction

Cultural NLP has grown as language technologies serve diverse users, but the field relies on varied cultural proxies without a shared conception of culture. The paper draws on sociocultural linguistics to clarify these theoretical and methodological challenges and motivate alternative goals.

1 Introduction

  • Language and culture are closely linked, and language technologies need to be robust to cultural differences across diverse user bases.Language is both an artifact of culture and a process through which culture is created.Source: §1 · p. 1
  • Cultural NLP builds or evaluates systems that know cultural facts and apply that knowledge appropriately in relevant situations.The work includes benchmarks, fine-tuning datasets, and systems that generate culturally relevant output (Adilazuarda et al., 2024; Liu et al., 2024b).Source: §1 · p. 1
  • The field’s varied proxies for cultural boundaries and objects point to a theoretical gap: a clearer, more unified conception of culture is needed.Proxies include nationality, religion, and ethnicity, as well as foods, values, and norms (Zhou et al., 2024a; Sorensen et al., 2024; Dwivedi et al., 2023).Source: §1 · p. 1
  • The paper uses sociocultural linguistics to clarify cultural NLP’s goals, examine indexicality in a case study, and propose paths forward, including localization as a framing.It argues that localization may better frame the goals of much current work than cultural competence or understanding.Source: §1 · p. 2

2 The goals of cultural NLP

Cultural NLP pursues four desiderata: responsiveness and discernment, alongside breadth and depth of cultural coverage. These aims form two approximate sets of trade-offs.

2 The goals of cultural NLP

  • Adaptive systems respond differently to specific cultural contexts rather than producing the same output across contexts.Work often evaluates whether systems generate appropriate responses to cultural contexts, including answers consistent with a target culture (Cao et al., 2024a; Huang and Yang, 2023).Source: §2 · p. 2
  • Discerning systems avoid reductive stereotypes while recognizing that users’ expectations for cultural adaptation vary by setting.Users may want systems to understand regional or ethnic dialects without generating them (Lucy et al., 2024; Blaschke et al., 2024).Source: §2 · p. 2
  • Inclusive systems aim for breadth by performing well across many cultures, often using nationality and surveys to scale coverage.Large-scale benchmarks and text-mining methods reflect this emphasis (Bhutani et al., 2024; Zhao et al., 2024; Ramezani and Xu, 2023; Fung et al., 2024; Nguyen et al., 2023).Source: §2 · pp. 2–3
  • Nuanced systems seek depth through granular, extensive cultural understanding, often trading broad coverage for locally rich knowledge.Some works use local informants and explore locally meaningful categories (Koto et al., 2024; Son et al., 2025; Li et al., 2024b; Dev et al., 2023).Source: §2 · p. 3
  • The four desiderata roughly form two trade-offs: cultural competence concerns how and when to respond, while coverage concerns how many cultures and cultural aspects are represented.The desiderata are not mutually exclusive.Source: §2 · p. 3

3 Recurring troubles

A survey of cultural NLP papers finds recurring limitations in coverage, dynamicity, and the definition of culture. The authors argue that these difficulties call for theoretical answers, while noting other concerns around data and evaluation.

3 Recurring troubles

  • 40% of surveyed papers cited limited cultural coverage, and 36% used nationality as a demographic proxy.Limitations describe national boundaries as lacking nuance because nations are politically defined and not culturally homogeneous.Source: §3 · p. 3
  • 12% of papers cited a lack of dynamicity: benchmarks are largely static despite culture being continually constructed through social negotiation.The surveyed works often lack temporal granularity (Ochs, 2009; Son et al., 2025; Keleg and Magdy, 2023; Jin et al., 2024; Li et al., 2024b).Source: §3 · p. 3
  • Roughly 37% of papers problematized a cultural proxy, with 28% specifically noting that proxies can miss intragroup variation.Many works focus on a small subset of cultural objects without a framework to unify them.Source: §3 · p. 3
  • The adaptation–discernment tension leaves some papers treating stereotypes as cultural knowledge and others collecting them to reduce stereotype generation.The former concern appears in 14% of papers, while other papers collect stereotypes to build systems that avoid generating them (Shi et al., 2024; Bhutani et al., 2024).Source: §3 · p. 3
  • Other reported concerns include English-language emphasis, multiple-choice knowledge tests, pretrained-model dataset construction, and conflating individual preferences with cultural ones.The authors say little progress has been made on rigorously addressing these limitations, which raise questions requiring theoretical answers.Source: §3 · p. 3

4 A sociocultural solution

The paper takes social context as a starting point for understanding culture and adopts sociocultural linguistics as its framework. Its concepts frame culture as emergent, layered, constructed through signs, relational, and necessarily incomplete.

4 A sociocultural solution

  • Cultural competence is framed through communicative conventions and a semiotically rich social space, making social context a starting point for understanding culture.The paper draws on linguistic anthropology and sociolinguistics, which study language and culture together.Source: §4 · p. 4

4.1 Text and context

  • Sociolinguistics moved from coarse social categories toward local dynamics and then to identity as performance through diverse semiotic resources.The paper adopts sociocultural linguistics as a convergent framework for analyzing language and other semiotic practices.Source: §4.1 · p. 4

4.2 A primer on sociocultural linguistics

  • Emergence treats culture as constituted through interaction, supporting dynamic representations and the possibility of inducing cultural categories latent in data.This view also allows an individual to inhabit multiple cultural roles.Source: §4.2 · p. 5
  • Positionality includes macro-level demographics, local distinctions, and context-specific stances and styles, rather than treating national identity as the only level.The framework also includes categories specific to a community or interaction.Source: §4.2 · p. 5
  • Indexicality explains culture’s construction through links between linguistic or other forms and social meaning.These links can shape ideologies, interactional stances, and overt references to identity.Source: §4.2 · p. 5
  • Relationality locates identity’s social meaning in relation to other identities and suggests encoding relations beyond similarity and difference.The framework also includes relations such as authentication–denaturalization and authorization–illegitimization.Source: §4.2 · p. 5
  • Partialness means accounts of culture are necessarily incomplete and situated, so researchers should examine their position and the assumptions encoded in systems.Identity at a given time may reflect intention, habit, perception, interactional context, and surrounding ideologies.Source: §4.2 · p. 5

5 Case study: culture from text

The case study distinguishes text about culture from culture as surrounding social context, then applies indexical theory to clarify what cultural-discourse datasets actually capture. It also shows that indexical meanings depend on context and can persist as their social basis changes.

5 Case study: culture from text

  • These systems extract mappings between beliefs and cultural groups from online documents, but such data excludes knowledge not subject to metapragmatic discussion.Indexicality clarifies the distinction between the text being studied and the surrounding social context treated as culture.Source: §5 · pp. 5–6, §5.1 · p. 6

5.1 Stereotypes all the way down

  • Cultural-discourse mining primarily captures stereotypes embedded in the data generators’ ideologies, not necessarily the cultures being discussed.The associations count as stereotypes because they are discussed, regardless of whether the speakers belong to the group they describe.Source: §5 · p. 5, §5.1 · p. 6
  • Higher-order cultural associations may remain salient after the lower-order identities they arose from have become less relevant.The Pittsburgh style of being unpretentious, hospitable, and sports-loving was linked to an immigrant-heritage working-class history that may not fit many current residents.Source: §5.1 · p. 6
  • Computational studies can also identify first-order social meaning using platform metadata such as geolocation or subreddit.These methods measure sociolinguistic variation and the social meanings it indexes, including locale or community.Source: §5.1 · p. 6

5.2 Indexical values are contextual

  • An utterance’s social meaning depends on its interactional context, so one style can function as accommodation or mockery.A “foreign speaker” style may even fulfill both roles, depending on the interactional frame.Source: §5.2 · p. 6

6 Paths forward

The paper identifies gaps between sociocultural theory’s account of culture and current computational methods, and outlines directions for culturally competent NLP. It argues that localization offers a more immediate framing by specifying an application, its audience, and expected behavior.

6 Paths forward

  • Sociocultural theory treats culture as complex and dynamic, while computational methods still leave gaps that NLP research could address.The paper presents cultural competence as a goal but notes that more immediate applications may operate under resource constraints.Source: §6 · p. 6, §6.1 · p. 7, §6.2 · p. 8

6.1 Culturally competent NLP

  • Datasets often capture static cultural artifacts, while studying interaction can reveal how cultural knowledge is shared, enforced, or contested.The exchange between Lupe and Patricia shows Spanish use and its contestation reinforcing its social importance as an index of Mexican identities.Source: §6.1 · p. 7
  • Culturally rich datasets should include social context and other semiotic resources, not only decontextualized text.Metadata or world state can situate interaction; multimodal resources such as gesture, prosody, fashion, and make-up can also convey social meaning.Source: §6.1 · p. 7
  • A system-generated rapport-building script can misfire when it implies personal experience the system does not have.The imagined grandmother’s movie-going history turns the script into a post-hoc rationalization rather than a way to motivate findings or establish rapport.Source: §6.1 · pp. 7–8

6.2 Localized NLP

  • Localization frames cultural NLP around a specified application, making relevant knowledge and evaluation situated and more tractable.It constrains the depth of cultural knowledge to domain-specific needs and focuses evaluation on performance in an application setting.Source: §6.2 · p. 8
  • Localization makes system coverage explicit by enumerating the audience and the cultural boundaries relevant to an application.For a healthcare chatbot, medical literacy may be a more actionable boundary than nationality; boundary choices should be explicit and transparent.Source: §6.2 · p. 8
  • Localization specifies expected system behavior in context, including when an application should adapt and when it should show cultural humility.A recipe app might offer culturally specific ingredient substitutions, while a healthcare app might avoid adaptation that stereotypes or stigmatizes.Source: §6.2 · p. 8

7 Conclusion

The paper argues that cultural NLP needs a theoretical account of culture, treating it as an enacted process rather than a static collection of facts. It presents sociocultural linguistics as a framework for this shift and argues that localization better frames many current goals.

7 Conclusion

  • Current cultural NLP lacks agreement on how to model culture and instead uses varied proxies for identity and indexical features.The paper presents this as evidence of a lack of theoretical alignment.Source: §7 · p. 9
  • Cultural competence requires a dynamic model of culture as a process, not a collection of trivia.The paper argues that sociocultural linguistics provides a powerful model, while methodological and theoretical gaps remain.Source: §7 · p. 9
  • The paper argues that many current cultural NLP works can be usefully reframed as localization, with situated, participatory system design and evaluation.Source: §7 · p. 9
  • As NLP has advanced in representing semantic meaning, sociocultural meaning has become a tractable and necessary object of study.Source: §7 · p. 9

8 Limitations

The paper limits its scope: it is not a survey of cultural NLP, and it focuses on sociocultural linguistics among the available theories of culture.

8 Limitations

  • The paper highlights representative works to support its theory-led argument, but is not intended as a survey of the field.Source: §8 · p. 9
  • Adilazuarda et al. (2024) and Liu et al. (2024b) are identified as overviews of recent cultural NLP work.Source: §8 · p. 9
  • The paper focuses on sociocultural linguistics because of its linguistically oriented approach to culture, while acknowledging that other theories exist.Source: §8 · p. 9

A List of surveyed works

The surveyed-work list spans cultural NLP topics including cultural knowledge, language understanding, norms, values, and evaluation.

A List of surveyed works

  • The listed works address cultural knowledge and language tasks, including time expressions, food-related knowledge, and culturally aware inference.Examples include Shwartz (2022), Huang and Yang (2023), and Zhou et al. (2024a).Source: A List of surveyed works · pp. 16–17
  • The cited titles concern socio-cultural norm discovery, pluralistic human values, rights and duties, and Korean bias benchmarking for question answering, covering norms, values, and bias.Source: A List of surveyed works · pp. 16–17
  • The list also includes work on multilingual and cross-cultural language systems, cultural adaptation, and community-driven resources.Examples include Multi-VALUE (Ziems et al., 2023b), Cultural Adaptation of Recipes (Cao et al., 2024b), and CultureBank (Shi et al., 2024).Source: A List of surveyed works · pp. 16–17

B Category key

The category key defines recurring labels for limitations and concerns reported in surveyed cultural NLP work, spanning proxies, coverage, data, and evaluation.

B Category key

  • Coverage, coarseness, multilingual, and proxy label limited scope, overly broad cultural boundaries, language coverage, and proxy choice, respectively.Source: B Category key · p. 17
  • Stereotype, culture is dynamic, context, and unmarked culture capture concerns about stereotypes, cultural change, missing situation, and assumed universal values.Source: B Category key · p. 17
  • Intrinsic and based on existing model identify evaluation limitations, while exploited annotators flags concerns about who benefits from preference data.Source: B Category key · pp. 17–18
  • Crowdsourced, annotator, subjectivity, and LLM-derived mark concerns about data sources and the subjectivity of cultural annotations.Source: B Category key · pp. 17–18
  • Pluralism, discerning, preference v morality, and individual distinguish value-alignment questions from treating individual values as cultural ones.Source: B Category key · pp. 17–18
Loading 2502.12057v2…