Cultural Proximity

• Cultural Proximity is a set of constructs defining how closely cultural artifacts, texts, and practices resemble each other based on domain-specific signals.
• Researchers measure it using varied frameworks, from human ratings and linguistic proxies to hyperbolic embeddings and Euclidean distances.
• Its applications span media studies, NLP translation, social network analysis, and urban mobility, offering practical insights for model alignment and policy design.

Cultural proximity is a family of analytic constructs used to describe how close cultures, cultural artifacts, texts, places, or model outputs are to one another, but its exact meaning varies by domain. In communication and media studies, it denotes audiences’ preference for culturally similar content because such content is easier to process, resonates with identity, and feels authentic; in translation and NLP, it is often operationalized as familiarity and resonance for a target readership; in social science and computational social research, it is measured through distance in value profiles, semantic positions, event distributions, mobility patterns, or visual concept representations (Luther et al., 13 Dec 2025, Zou et al., 16 Sep 2025, Wu et al., 2018). Across recent work, cultural proximity is therefore not a single ontology-backed quantity but a domain-specific relation between cultural forms, actors, or populations.

1. Conceptual scope and distinctions

Cultural proximity is commonly defined as a form of closeness that is not exhausted by literal semantic equivalence or by physical distance. In wine-review adaptation, it is “the degree to which the generated review uses familiar terms and expressions that resonate with the target-culture reader,” and the criterion asks whether an adapted review “feels native” through familiar flavor descriptors, metaphors, and stylistic choices (Zou et al., 16 Sep 2025). In LLM alignment work, it is the descriptive closeness between a model’s inferred cultural value profile and a target country’s profile, while in social-semantic analysis it is the tendency for socially connected actors to occupy nearby positions in a shared cultural space, with closeness reflected by smaller hyperbolic distances and tighter angular alignment on a Poincaré disk (Luther et al., 13 Dec 2025, Wu et al., 2018).

Several adjacent concepts are explicitly separated from cultural proximity in the literature. In wine adaptation, Cultural Proximity differs from Cultural Neutrality, which concerns avoiding terms with negative or awkward connotations in the target culture, and from Cultural Genuineness, which concerns preserving the authentic sensory quality of the original descriptor without distorting meaning (Zou et al., 16 Sep 2025). In hyperbolic social-semantic models, proximity differs from hierarchy, encoded by radial position, and from polarization, encoded by large angular separations with enduring mutual awareness (Wu et al., 2018). In cross-lingual transfer, pragmatic cultural proximity is distinguished from typological and genealogical similarity; the claim is that shared communicative norms, figurative conventions, and emotion semantics can matter more than syntax or language family for pragmatically motivated tasks such as sentiment analysis (Sun et al., 2020).

Recent multimodal work also distinguishes cultural proximity from geography in a strict sense. ArtiFact treats it pragmatically through geographic adjacency, continental co-membership, and temporal-stylistic neighborhood, but explicitly does not provide a formal mathematical definition or ontology-backed model (Duarte et al., 8 Jun 2026). CUNIT uses a coarse Eastern-versus-Western grouping and country-pair similarity heatmaps rather than explicit geographic-distance metrics, and its results indicate that geo-cultural proximity has only a weak influence on model performance in identifying cross-cultural concept associations (Li et al., 2024). This suggests that cultural proximity may incorporate geography, but is not reducible to it.

2. Measurement frameworks

Representative operationalizations differ in the object being compared, the signal used, and the role assigned to human judgment, corpora, or survey instruments.

Setting	Object compared	Proximity signal
Wine-review adaptation	Review and target readership	Seven-point human rating of familiarity and resonance (Zou et al., 16 Sep 2025)
Cross-lingual pragmatics	Source and target languages	LCR, LTQ, and ESD features (Sun et al., 2020)
LLM cultural alignment	Model and country	Distance over Hofstede dimensions (Luther et al., 13 Dec 2025)
Social-semantic systems	Actors or ideas	Hyperbolic distance and angular span (Wu et al., 2018)
Visual concept representation	Country and country	Distances between sketch-based concept signatures (Pera et al., 8 Jul 2026)

Human-evaluation frameworks often treat cultural proximity as a scalar judgment. In wine-review adaptation, annotators rate Cultural Proximity on a seven-point Likert-type scale from 7 = excellent to 1 = nonsense, and aggregate scores by

$$\bar{s}_i=\frac{1}{M}\sum_{j=1}^{M}s_{ij}, \qquad S=\frac{1}{N}\sum_{i=1}^{N}\bar{s}_i.$$

The same study uses Jaccard similarity to quantify cross-cultural descriptor overlap,

$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$

showing significant cross-cultural gaps at the levels of Exact Aromas, Subfamilies, and Families (Zou et al., 16 Sep 2025).

Feature-based linguistic approaches operationalize proximity through measurable proxies for pragmatics. The cross-lingual transfer literature defines pronoun- and verb-based context-level ratios,

$$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$

a Literal Translation Quality feature based on average hit ratios over MWEs, and Emotion Semantics Distance,

$$ESD(l_{tf},l_{tg})=\frac{\sum_{e\in E}\cos(\mathbf{v}_{tf,e},\mathbf{v}_{tg,e})}{|E|}.$$

These are not generic similarity measures; they are proposed as cultural proxies grounded in context-level communication, figurative language, and emotion semantics (Sun et al., 2020).

Geometric and network approaches model proximity as distance in an embedding space. In the Poincaré-disk formulation, the hyperbolic distance between two points is

$$d(u,v)=\operatorname{arcosh}\left(1+\frac{2\|u-v\|^2}{(1-\|u\|^2)(1-\|v\|^2)}\right),$$

while angular diversity can be summarized by

$\Theta=\max_i\phi_i-\min_i\phi_i,$

with smaller $\Theta$ indicating contraction (Wu et al., 2018). Event-based country clustering instead uses Euclidean distance on rank vectors of topic prevalence, whereas neighborhood-mobility models use cosine similarity between amenity vectors and between scene vectors,

$$S^{amen}_{ij}=\frac{f_i\cdot f_j}{\|f_i\|\|f_j\|}, \qquad S^{scene}_{ij}=\frac{n_i\cdot n_j}{\|n_i\|\|n_j\|}.$$

These formulations treat proximity as similarity in cultural profile rather than semantic equivalence (Tama et al., 2023, Silva et al., 12 Dec 2025).

Survey-based cultural-alignment work uses national value profiles. One line computes Euclidean cultural distance over Hofstede’s six dimensions,

$$CD(C_1,C_2)=\sqrt{(pdi_{C_1}-pdi_{C_2})^2+\cdots+(ivr_{C_1}-ivr_{C_2})^2},$$

with “culturally close” pairs defined by lower-quartile thresholds in the full pairwise distribution (Evan et al., 11 Apr 2025). LLM alignment studies instead use the sum of absolute differences across the same six dimensions,

$$D_{\text{model},\text{country}}=\sum_{i\in\{PDI,IDV,MAS,UAI,LTO,IVR\}}\left|s_i^{\text{model}}-s_i^{\text{country}}\right|,$$

or Euclidean distance on the two-dimensional Inglehart–Welzel cultural map after projecting model responses into IVS-derived coordinates (Luther et al., 13 Dec 2025, Tao et al., 2023).

Multimodal concept research introduces a different construction: country-specific visual concept signatures based on odds ratios over sketch clusters. For country $J(A,B)=\frac{|A\cap B|}{|A\cup B|},$0 and cluster $J(A,B)=\frac{|A\cap B|}{|A\cup B|},$1,

$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$2

and global sketch-based proximity is then computed from Euclidean distances between z-scored log-odds vectors, averaged across concepts (Pera et al., 8 Jul 2026).

3. Language, translation, and LLM-mediated cultural resonance

In NLP, cultural proximity frequently appears where literal fidelity is insufficient. The wine-review benchmark shows that standard machine translation and state-of-the-art LLMs still struggle to capture cultural nuance across Chinese and English tasting discourse. Automatic metrics correlate weakly with Cultural Proximity; for Chinese$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$3English, correlations with BLEU, METEOR, BERTScore, and BEER are near zero or negative, and for English$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$4Chinese only BERTScore shows a small but significant positive correlation with Cultural Proximity, $J(A,B)=\frac{|A\cap B|}{|A\cup B|},$5 (Zou et al., 16 Sep 2025). Quantitatively, Mistral reaches the highest Cultural Proximity for Chinese$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$6English at 6.7, while Qwen2.5 is highest for English$J(A,B)=\frac{|A\cap B|}{|A\cup B|},$7Chinese at 6.0, close to the human score of 5.8; for ChatGPT, Self-Explanation raises Cultural Proximity to 5.75 but tends to trade off faithfulness (Zou et al., 16 Sep 2025).

The same domain makes clear what high-proximity adaptation looks like: substituting culturally familiar descriptors while maintaining sensory plausibility. Examples include adapting raspberry to blueberry for broader familiarity in China, contextualizing black currant with references such as loquat paste or hawthorn cake, and rendering “earthy” as 泥土气息 rather than 土味 (Zou et al., 16 Sep 2025). Low-proximity failures include literal carry-over of rare botanicals, misinterpretation of multiword descriptors such as “Mediterranean scrub,” register mismatches, and domain-term errors such as translating “nose” as a literal body part (Zou et al., 16 Sep 2025).

Cross-lingual transfer studies generalize this beyond wine. For sentiment analysis, adding pragmatic cultural features to LangRank raises MAP from 71.3 to 76.0 and NDCG from 86.5 to 90.9, while for dependency parsing typology remains the stronger guide, indicating that the relevance of cultural proximity is task-dependent rather than universal (Sun et al., 2020). In Offensive Language Detection, Hofstede-based cultural features and offensive-word distance improve source-language ranking for transfer; for example, adding cultural features to MTVec raises MAP from 51.67 to 70.28 and NDCG from 68.23 to 76.92, and the combined P+C+O setting yields MAP 76.67 and NDCG 79.61 (Zhou et al., 2023).

Benchmarks of LLM cultural reasoning produce a more mixed picture. CPopQA shows that GPT-3.5 can rank long-tail cultural concepts by popularity and exhibits potential to identify geo-cultural proximity across continents, but performance degrades as ranking complexity increases and on under-represented countries (Jiang et al., 2023). CUNIT, by contrast, finds that geo-cultural proximity exerts only a weak influence on model performance in identifying highly associated cross-cultural concept pairs; explicit feature scaffolds help more, especially for geographically distant cultures (Li et al., 2024). A plausible implication is that proximity effects are easier to recover when the task is corpus-statistical and weakly structured than when it requires feature-level analogical reasoning about culturally specific concepts.

4. Social, spatial, and multimodal forms of proximity

Outside language technology, cultural proximity is often studied as a structure in social or geographic space. In the APS collaboration study, social and semantic networks are embedded in comparable two-dimensional hyperbolic spaces, and denser collaboration is associated with contraction of topic diversity. Pairwise social and cultural “diverging speeds” correlate at $J(A,B)=\frac{|A\cap B|}{|A\cup B|},$8, current social distance predicts next-year cultural distance in 67% of significant pairwise regressions, and “institution vs. world” analyses raise that share to 89% (Wu et al., 2018). Cultural proximity here is not merely convergence; polarization can also coexist with contraction, because attention concentrates into opposed sectors even as overall variety falls (Wu et al., 2018).

Migration research operationalizes cultural proximity with Hofstede-based Euclidean distance and then stratifies origin–destination pairs into culturally close, mid-distant, and distant groups. Using quartiles over 93 countries, “culturally close” means $J(A,B)=\frac{|A\cap B|}{|A\cup B|},$9 and “culturally distant” means $$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$0 (Evan et al., 11 Apr 2025). The results indicate that approximately half of OECD countries demonstrate a statistically significant bias toward accepting culturally close migrants, that migrants from wealthier countries tend to select culturally similar destinations, and that about two-thirds of OECD countries show positive alignment between cultural similarity and geographic proximity, with notable exceptions such as New Zealand and Australia (Evan et al., 11 Apr 2025).

Event-based and urban-mobility analyses shift the unit from countries to places. Meetup-based clustering defines proximity between countries as similarity in offline event-topic distributions and finds that national economic status explains 44.6 percent of the variance in total event count, while cultural characteristics such as Individualism and Long-Term Orientation explain 32.8 percent of the variance in topic categories (Tama et al., 2023). At the neighborhood level, similarity in amenity mix and cultural scenes acts as a “soft infrastructure” of mobility: in the United States, a one-standard-deviation increase in amenity similarity is associated with a 14.2% increase in co-visitation and a one-standard-deviation increase in scene similarity with a 5.0% increase; in Canada, the analogous increases for residential mobility are 17.7% and 8.2%, with a positive interaction between amenity and scene alignment (Silva et al., 12 Dec 2025).

Cultural-heritage and visual-concept datasets demonstrate proximity under multimodal ambiguity. ArtiFact operationalizes cultural proximity through adjacency-based culture and geographic manipulations, statistical groupings across time and place, and normalization of culture versus location fields; semantic query systems then struggle precisely where neighboring cultures share materials and styles, as in “British alabaster objects” and “swords from China” (Duarte et al., 8 Jun 2026). The sketch-based concept study takes a more direct route by comparing how countries visually realize the same concepts. It reports that cross-cultural similarities derived from sketches align 45% more closely with established cultural distances than text-based measures do, and its image-derived network reveals coherent regional communities such as English-speaking countries, South America, Europe, post-Soviet Eurasia, Africa and the Middle East, and Asia (Pera et al., 8 Jul 2026).

5. Limits, controversies, and recurrent failure modes

A recurrent limitation is that cultural proximity is easier to invoke than to formalize. ArtiFact explicitly lacks a formal ontology or authority file for cultural or stylistic lineages and instead uses heuristics such as neighboring countries, continent groupings, and statistical co-occurrence (Duarte et al., 8 Jun 2026). The wine-review benchmark provides detailed annotator guidance and a 40-sample pilot but does not report inter-annotator agreement values for the main study, and it also notes that the Chinese portion of the corpus is dominated by a few critics and that subjectivity is inherent in tasting notes (Zou et al., 16 Sep 2025). These are not minor details: they indicate that many current proximity measures remain operational rather than canonical.

Another debate concerns reducibility. The cross-lingual pragmatics work reports only modest correlations between its pragmatic features and geography—ESD at $$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$1, LCR-pron at $$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$2, LCR-verb at $$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$3, and LTQ at $$LCR\text{-pron}(l_{tf},l_{tg})=\frac{ptr(l_{tg})}{ptr(l_{tf})}, \qquad LCR\text{-verb}(l_{tf},l_{tg})=\frac{vtr(l_{tg})}{vtr(l_{tf})},$$4—and argues that pragmatic cultural similarity is not reducible to spatial proximity (Sun et al., 2020). CUNIT likewise finds only weak performance effects from geo-cultural proximity and recommends explicit cultural feature grounding rather than assuming that regional closeness will suffice (Li et al., 2024). This suggests that simple geographic or genealogical shortcuts often fail for pragmatically loaded or conceptually fine-grained tasks.

Survey-based alignment frameworks introduce a different controversy: they are transparent and comparable, but they risk flattening within-country heterogeneity. LLM studies based on VSM13 or the Inglehart–Welzel map repeatedly warn that prompting a model to act like “an average person” from a country can reinforce monolithic views and overlook subcultures and diversity (Luther et al., 13 Dec 2025, Tao et al., 2023). Empirically, these same studies report strong Western or US defaults: five of eight models in one VSM13 study were closest to the United States by default, cultural prompting improved alignment for 71–81% of countries in IVS-based evaluation for recent GPT models, but France prompting sometimes moved models closer to US values than to France, and Japan and China remained difficult targets (Luther et al., 13 Dec 2025, Tao et al., 2023, Luther et al., 10 Dec 2025).

Proximity can also have dual effects rather than uniformly beneficial ones. In misinformation detection, South Africans were better at recognizing true South African news than other nationalities, with 40% versus 52% deviation from the ideal rating, but worse at identifying fake news, with 62% versus 55% deviation; the study associates this asymmetry with higher trust and greater reliance on contextual knowledge (Schlippe et al., 21 Nov 2025). In the APS collaboration study, increased social connection raises mutual information and reduces joint entropy, producing cultural contraction even when the resulting structure is polarized rather than consensual (Wu et al., 2018). A plausible implication is that cultural proximity can improve fit, fluency, or verification in some settings while simultaneously narrowing diversity or increasing vulnerability to plausible, locally resonant distortions.

6. Research directions and practical significance

Across domains, recent work recommends moving from coarse proxies toward culture-aware, task-specific systems. In translation and localization, recommended strategies include culture-aware fine-tuning and retrieval with merged aroma-wheel lexicons, detection-and-explanation pipelines that flag culturally opaque descriptors, multi-reference evaluation, larger multilingual corpora, multimodal grounding, and user-in-the-loop feedback (Zou et al., 16 Sep 2025). In cross-cultural concept matching, the proposed agenda includes richer cultural feature ontologies, explicit proximity signals such as geographic distance or historical ties, and broader coverage beyond clothing and food (Li et al., 2024).

For cultural-heritage and multimodal data management, the immediate need is domain knowledge and proximity-aware reasoning. ArtiFact is positioned as a challenging benchmark precisely because current systems cannot reliably distinguish neighboring cultures, resolve historically contingent terminology, or account for how object appearances evolve across time and geography without more explicit cultural-historical grounding (Duarte et al., 8 Jun 2026). The sketch-based work points in a complementary direction: multimodal measurement can reveal conceptual structure that language compresses, especially for concepts with strong perceptual or haptic content (Pera et al., 8 Jul 2026).

LLM evaluation research proposes standardized survey pipelines, per-dimension monitoring, and locale-aware prompting. Recommended practices include eliciting multiple stochastic runs, computing transparent distance metrics against country benchmarks, tracking dimension-wise errors rather than only total scores, and adding local-language prompting because English-only prompting can itself bias responses toward Western or US norms (Luther et al., 13 Dec 2025, Luther et al., 10 Dec 2025). Cultural prompting is presented as a practical control strategy, but the same literature emphasizes ongoing auditing rather than one-time correction (Tao et al., 2023).

In social and policy applications, proximity remains consequential even when not directly manipulable. Migration research recommends integration strategies tailored to cultural gaps, labor-market matching that accounts for communication and norm frictions, and future models that combine cultural distance with economics, policy, and historical ties (Evan et al., 11 Apr 2025). Urban-mobility research recommends aligning functional and symbolic place qualities to reduce isolation and strengthen connectivity, since neighborhoods with similar cultural styles and amenities are more strongly connected even after controlling for race, income, education, politics, housing costs, and distance (Silva et al., 12 Dec 2025).

Taken together, these lines of work indicate that cultural proximity is best understood not as a universal scalar but as a family of socially grounded relations between actors, texts, values, places, and representations. This suggests that rigorous analysis requires explicit attention to domain, modality, and evaluation regime: familiarity and resonance in translation, pragmatic similarity in language transfer, value-profile distance in survey alignment, heuristic neighborhood in heritage data, and embodied exemplar similarity in visual concepts are all genuine but non-identical instances of cultural proximity.