Plurification in/of language technology -- The integration of culture in next-generation AI

Technical Summary: Plurification in/of Language Technology

Problem Statement
The paper addresses the critical challenge of operationalizing 'culture' within Natural Language Processing (NLP) and Language Technology (LT). Current approaches often treat culture as a static, measurable variable or attempt to address it merely by adding more examples of "other cultures" to training data. The authors argue this is insufficient because culture is dynamic, internally diverse, and deeply entangled with power, identity, and context. Existing NLP systems frequently fail to model the social norms and background knowledge that shape interpretation, often reproducing dominant discourses while underrepresenting minority voices, local knowledge, and culturally specific meanings. The core problem is that cultural alignment cannot be achieved through superficial technical adaptation (e.g., prompt design or post-hoc moderation) but requires a fundamental shift toward plural epistemologies—allowing multiple, locally grounded ways of knowing to coexist within technological design.

Methodology
The paper employs a qualitative, case-based approach grounded in the anthropology of technology. It utilizes a socio-technical model of language technology design known as the "five layers of technological activity" (Koch et al., 2024) to systematically analyze and systematize existing approaches to culture in NLP. This model organizes the analysis across five distinct layers:

1. Technological Core: Software, model architectures, training procedures, and decoding strategies.
2. Knowledge Domain: Corpora, linguistic resources, and symbolic categories used for training.
3. Knowledge Ecologies: Broader social, normative, and pragmatic environments (e.g., local knowledge systems, legal frameworks).
4. Political Economies: Data extraction practices, annotation labor, funding, and power asymmetries.
5. Contexts of Use: The actual settings and communities where technologies are deployed and interpreted.

The authors select specific examples from NLP research that demonstrate attempts to integrate cultural dimensions across these layers. The selection criteria focus on whether an intervention broadens the count of languages/knowledge, challenges universalizing assumptions, involves local expertise, and whether it is an embedded feature or a superficial output adaptation.

Key Contributions and Analysis
The paper synthesizes research to demonstrate how "plurification" (the integration of multiple epistemologies) can occur across the five layers, moving beyond shallow cultural adaptation:

• Layer 1 (Technological Core): The paper highlights the work on the Universal Knowledge Core (UKC) (Giunchiglia et al.), which proposes a language-neutral semantic layer. Instead of centering English as a pivot language, the UKC represents meaning at the concept level, allowing multiple languages to map their lexical forms onto a shared structure. This supports many-to-many multilingual relations without linguistic homogenization.
• Layer 2 (Knowledge Domain): The analysis draws on Steven Bird's work connecting NLP to language documentation, emphasizing that datasets are constructed artifacts with social histories. It advocates for data statements and formal annotation frameworks to improve the availability and interoperability of resources for low-resource languages. Giunchiglia's work on ontologies is also cited to show how structured conceptual representations can align heterogeneous knowledge systems without forcing them into a single normative schema.
• Layer 3 (Knowledge Ecologies): The paper examines normative frameworks (e.g., African Union AI frameworks, UNESCO guidelines) that advocate for diversity-aware AI. However, it critically notes that current corporate-driven frameworks often remain superficial and profit-oriented, creating an "AI governance crisis" where local and regional knowledge is not adequately integrated.
• Layer 4 (Political Economies): The authors analyze the extractive nature of current LT, where communities are treated as data sources rather than co-designers. Citing Nekoto et al. and Bird, the paper argues for participatory machine translation and sustainable language development. It emphasizes the need for fair compensation, ownership, and benefit-sharing for the labor involved in data collection and annotation, suggesting a shift from benchmark optimization to community benefit.
• Layer 5 (Contexts of Use): The paper argues that cultural appropriateness is situated and cannot be validated solely by technical metrics. It calls for treating speech communities as epistemic authorities. "Plurification" at this level involves creating "third spaces" where local frameworks structure model goals and evaluation, ensuring systems are socially intelligible and not extractive.

Results
The analysis reveals that while NLP research has made progress toward culturally sensitive systems (e.g., dialect modeling, multilingual data), most interventions remain partial and focused on the output or representation layers. Deeper questions regarding power, governance, and social context remain unresolved. The paper finds that:

• Culture is not a stable variable but a dynamic socio-technical phenomenon.
• Technical adaptation alone (e.g., adding more data) fails to address the root causes of cultural misalignment.
• True operationalization of culture requires a reflexive approach that navigates the limits of computational formalization.
• Current approaches often fail to track cultural change over time or account for the high costs of doing so.

Significance and Claims
The paper claims that operationalizing culture in NLP requires a "plural socio-technical approach" rather than a purely technical fix. Its significance lies in reframing cultural diversity not as a data problem to be solved by more examples, but as a design orientation that must permeate all layers of technological activity. The authors posit that:

• Plurification is an ongoing process: It is not a final state but a continuous design orientation that redistributes epistemic authority toward speech communities.
• Socio-technical integration is essential: Culture-aware NLP must account for the interplay between algorithms, data, political economies, and local contexts.
• Limits of Formalization: The dynamic and multi-faceted nature of socio-cultural contexts means that culture can only be partially formalized; the system must remain responsive to disagreement and context-specific validity.

Ultimately, the paper argues that for language technology to be truly inclusive, it must move beyond the "imaginary of a super-intelligence" and embrace a pluriversal design that acknowledges multiple, locally grounded ways of knowing.