Liu, Chen Cecilia, Iryna Gurevych, and Anna Korhonen. 2025. “Culturally Aware and Adapted NLP: A Taxonomy and a Survey of the State of the Art.” Transactions of the Association for Computational Linguistics 13 (July): 652–89. https://doi.org/10.1162/tacl_a_00760.

Notes

• Taxonomy
  • Ideational
    • Concepts
      • basic concepts that are configured differently based on culture
      • culturally unique concepts
    • knowledge
    • Values
      • Biases
      • Hate
      • Other perceptions
    • Norms and morals
    • Artifacts (like art & poetry)
  • linguistic elements
    • Dialects (regionolects, sociolects, etc)
    • Styles, Registers, Genres
  • Social Elements
    • relationship
    • Context
    • Communicative goals
      • Different communication style based on communicative goals
    • Demographics
      The characterisation of 'NLP-system' here seems to be limited to just individual or consumer based applications. It is not taking domain or institutional contexts into consideration that further significantly expands the granularity of the localization and contextualization of the systems.

Culturally aware resource acquisition

how to make a system-culturally aware

• Manual (Incorporating Native Speakers, Communities, and Experts)
• Automatic (models and pipelines)
• Semi -automatic (Structured Resources, Model-in-the-loop)
  Source types
• New (ex: creating dataset from scratch)
• Culturally-adapted (adapting an existing dataset for a culture)

Creating culturally adapted models

• In-content adaptation
  • behavior tuning by prompts or in-context examples
  • RAG
• In-weight Adaptation
  • Data Augmentation
  • Continual Pre-training
  • Integrating cultural dimension vectors
  • parameter -efficient adaptations
  • Feedback learning

In-text annotations

"Hershcovich et al. (2022) proposed a simple taxonomy derived from the interaction be- tween language and culture that captures broad elements of culture (linguistic form and style, objectives and values, common ground, and about- ness)" (Page 652)

"language is an essential component of culture in NLP" (Page 652)

"Culture encompasses the collective ideas, shared language, and social practices that emerge from and evolve through human social interactions within a society" (Page 653)

"2 The Taxonomy" (Page 653)

"three main branches: ideational, linguistic, and social." (Page 653)

"The ideational branch (§3.1; Murdock, 1940; Briggle and Mitcham, 2012a) encompasses the non-material aspects of culture that constitute a way of life, such as values or knowledge. The linguistic branch (§3.2) focuses on cultural variations in language and linguistic forms, bridging the ideational and social elements of culture. The social branch (§3.3) covers key factors in social interaction and communication, such as relationships or commu- nicative goals.5 Here, we define each element based on existing research and relating to example tasks in the NLP context. We then provide details and examples from th" (Page 653)

"includes variations of languages in a Dialects: systematic way (Fromkin et al., 1998; Trudgill, 2000; Wardhaugh and Fuller, 2021; such as di- alects continuum, regionalects, sociolects, etc.)" (Page 654)

"y, integrating knowledge bases (KB) with models enhances cultural awareness (Bhatia and Shwartz, 2023) and supports culturally relevant synthetic data generation (Kim et al., 2023)" (Page 656)

"Many recent studies on evaluation (Johnson et al., 2022; Ramezani and Xu, 2023; Cao et al., 2023; Durmus et al., 2024; Santurkar et al., 2023; Masoud et al., 2025; Havaldar et al., 2023b; Wang et al., 2024c, inter alia) show that LLMs align better with values of WEIRD (Western, Educated, Industrialized, Rich and Democratic" (Page 656)

"In NLP, linguistic context could be the sur- rounding text." (Page 658)

"The extra-linguistic context can be situational (setting or location where communication occurs; e.g., at school, in a hospital), historical (past events; e.g., colonization, that change cultural values or language use, like in Hong Kong) or non-verbal (e.g., hand gesture, tone of voice). Each type shapes and reflects culture." (Page 658)

"Since different cultures reflect different values, there is a need to create models that embody pluralistic cultural values with flexible align- ment capabilities" (Page 660)

"RLHF fine-tunes LMs with feedback by fitting a reward model with human preferences, and then training a rein- forcement learning-based policy to maximize the learned reward. DPO avoids RL training by us- ing a simpler supervised learning objective for an implicit reward model" (Page 661)

"Evolving Culture: Culture evolves gradually (Boyd and Richerson, 1988; Whiten et al., 2011), yet there have been few discussions on how to model and adapt to evolving culture. Future research should focus on methods that address the dynamic nature of culture. One potential approach is the use of retrieval-augmented systems to integrate evolving information (§5.1), which ensures models’ relevance to cultural shifts over time" (Page 663)

"Adaptation in the Social Context: As a key mo- tivation of this paper, culture emerges from and is shaped by social interactions among humans within a society (§1). However, an important question remains unanswered in the existing literature: Should the cultural adaptation of mod- els occur within a situated social context and structure? Exploring this could present new av- enues for interdisciplinary research (e.g., with human-machine collaboration, social psychology, or anthropology)" (Page 663)