Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages

Weiran Wang; Wenjuan Qin; Haitao Liu

cs.CLOct 22, 2026

Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages

Authors: Weiran Wang, Wenjuan Qin, Haitao Liu

Organizations: Fudan University, Shanghai 200433, China

Abstract

Do people who speak different languages organize concepts in similar ways? This question lies at the heart of the Sapir--Whorf hypothesis and debates about linguistic relativity. Shared basic meanings, however, need not imply shared relations among them. We compared relations among 196 basic concepts across 81 languages, using word embeddings from two multilingual models to quantify conceptual structure expressed in language. Cross-language correspondence was greater for sociocultural meanings than for meanings concerning nature and motion, a contrast recurring across models and classifications. Within the sociocultural domain, relational differences increased with geographic separation; less closely related languages differed more across all domains. These findings suggest that universality and relativity are interwoven across conceptual domains. They inform an account in which shared cognitive resources interact with linguistic and local experience, and predict greater cross-language correspondence in human semantic judgments for sociocultural than for nature and motion meanings.

Figures & tables

Domain	$n$	Included fields
Biological Core	67	Body, perception, bodily processes, people, kinship, personal pronouns
Nature & Motion	49	Nonhuman living kinds, environments, movement, physical change
Socio-Cultural	59	Properties, evaluation, color, time, actions, material culture, numerals
Grammar & Function	21	Demonstratives, interrogatives, quantifiers, relational/function items

Table 1: Operational domains in the primary grouping. Counts refer to the 196 retained concepts; complete item assignments are provided in the Supplemental Material.

Figure 1: Cross-language variation in lexical relations. Multilingual BERT (mBERT) results for 81 languages and 196 concepts. The panels show (a) a descriptive principal-component projection of within-language centered vectors, colored by domain; (b) global representational similarity analysis (RSA) dissimilarities for 3,240 language pairs; and (c) domain distributions. The projection does not establish conceptual modules. Lower dissimilarity indicates more similar ordering of within-language relations; global and domain sets were centered separately.

Figure 2: Domain patterns across encoders and concept groupings. Points show raw mean representational similarity analysis (RSA) dissimilarities on matched language and concept support. Colors identify encoders; solid lines/filled markers indicate the primary grouping and dashed lines/open markers the reference-style grouping. Error bars show the range of means after omitting each language in turn, not confidence or credible intervals. Adjusted estimates and credible intervals are reported in the text and Supplemental Material.

Figure 3: Geographic and language-relatedness associations. All panels show multilingual BERT (mBERT). Total domain slopes and 95% credible intervals are shown in (a); filled circles indicate the primary grouping and open squares the reference-style grouping. Panels (b) and (c) show raw representational similarity analysis (RSA) dyad outcomes and descriptive least-squares lines, distinct from the conditional Bayesian estimates in (a). Genealogical distance is the word-form-based relatedness proxy defined in Method. Complete estimates for both encoders are in Supplemental Tables S15 and S16 .

Figure 4: Selected language comparisons illustrate lexical variation. Multilingual BERT (mBERT) examples selected for large directional gaps. The locations of the two triplets are shown in (a). Grammar & Function items in (b) compare Dutch with Afrikaans and Basque; Socio-Cultural items in (c) compare Icelandic with Marathi and Scots. Scores are item-level cosine similarities between language-centered vectors, not domain representational similarity analysis (RSA). Filled markers indicate the comparison with lower genealogical distance and open markers the geographically closer comparison. These selected examples are descriptive.

Domain	Semantic focus	Representative comparative and cognitive literature	Input / complete
Biological Core	Body, person, kinship, perception, vital processes, and primary-mapping personal pronouns	Core knowledge, embodied structure, kinship systems, and recurrent lexical organization ( Spelke and Kinzler, 2007 ; Kemp and Regier, 2012 ; Pagel, 2009 )	68 / 67
Nature & Motion	Nonhuman living kinds, physical environments, movement, and physical change	Frames of reference, motion descriptions, landscape language, and spatial cognition ( Levinson, 2003 ; Burenhult and Levinson, 2008 ; Talmy, 1985 ; Levinson and Wilkins, 2006 )	53 / 49
Socio-Cultural	Properties, evaluation, colour, time, actions, material culture, and numerals	Artifact naming, cultural categorization, emotion, odor, and communicative coordination ( Malt et al., 1999 ; Malt and Majid, 2013 ; Jackson et al., 2019 ; Majid and Burenhult, 2014 ; Thompson et al., 2020 )	60 / 59
Grammar & Function	Demonstratives, interrogatives, relations, quantifiers, and other function items	Grammaticalization, discourse functions, usage-based structure, and language–cognition interaction ( Beckner et al., 2009 ; Christiansen and Chater, 2008 ; Croft, 2001 ; Bybee, 2010 ; Levinson and Wilkins, 2006 )	26 / 21

Table S1: Literature motivation and support for the four operational domains. Counts refer first to the 207-item input and then to the 196-concept complete-case support.

Term	Mean	95% CrI	Interpretation
Primary; Model (XLM-R-large)	$0.0311$	$[0.0277,0.0345]$	Model main effect relative to mBERT Biological Core.
Primary; Model $\times$ Nature & Motion	$0.0211$	$[0.0162,0.0260]$	Model-by-domain increment.
Primary; Model $\times$ Socio-Cultural	$0.0167$	$[0.0119,0.0213]$	Model-by-domain increment.
Primary; Model $\times$ Grammar & Function	$0.0435$	$[0.0387,0.0483]$	Model-by-domain increment.
Primary; Model $\times$ Geography	$0.0060$	$[0.0023,0.0096]$	Difference in baseline geography slopes.
Primary; Model $\times$ ASJP	$-0.0052$	$[-0.0088,-0.0015]$	Difference in baseline ASJP slopes.

Table S2: Formal pooled cross-model posterior contrasts. The mBERT Biological Core intercept is the reference. Intervals are equal-tail 95% credible intervals from the full-support pooled model.

Taxonomy	Full estimate	2.5%	97.5%
Primary	0.0044	0.0012	0.0074
Reference-style taxonomy	$-0.0029$	$-0.0062$	0.0002

Table S3: Paired language-dyad bootstrap for the cross-model endpoint contrast. The estimand is

(\text{XLM-R-large}-\text{mBERT})

applied to the Nature & Motion minus Socio-Cultural contrast. Both models and all four domains are resampled together within each of 2,000 complete-dyad replicates.

Taxonomy	All	Biological	Nature	Socio-Cultural	Grammar
Primary	.9321	.9315	.9767	.9220	.9671
External	.9321	.9529	.9729	.9383	.9632

Table S4: Full-valid XLM-R coverage check. Entries are raw RSA dissimilarity means for the same 81-language, 196-concept, 3,240-dyad support; no Bayesian intervals are assigned to this coverage description.

Analysis	Taxonomy	Full (%)	S $<$ N (%)	S min. (%)	N max. (%)
Dyad bootstrap	Primary	100.0	100.0	100.0	100.0
Dyad bootstrap	Reference-style	99.75	100.0	100.0	99.75
Equal-size concepts	Primary	33.6	89.6	65.4	70.8
Equal-size concepts	Reference-style	35.4	95.8	69.4	54.4

Table S5: Dependence- and domain-size-sensitivity results. Percentages are proportions of resamples or equal-size concept subsamples. The dyad bootstrap retains all four domain outcomes within each sampled dyad.

ID	Concept	Language	Status	Domain
201	at	Japanese	all-zero-vector	Grammar & Function
202	in	Japanese	all-zero-vector	Grammar & Function
203	with	Japanese	all-zero-vector	Grammar & Function
204	and	Japanese	all-zero-vector	Grammar & Function
206	because	Japanese	all-zero-vector	Grammar & Function
58	bark	Marathi	all-zero-vector	Nature & Motion

Table S6: Invalid concept–language observations identified before RSA. All 12 were all-zero vectors in the extraction and were treated as missing.

ID	Concept	Primary domain	Valid	Invalid observations
2	you	Biological Core	80	Vietnamese:all-zero-vector
58	bark	Nature & Motion	79	Marathi:all-zero-vector; Nepali:all-zero-vector
125	stand	Nature & Motion	80	Occitan:all-zero-vector
148	moon	Nature & Motion	80	Thai:all-zero-vector
149	star	Nature & Motion	80	Thai:all-zero-vector
192	dull	Socio-Cultural	80	Occitan:all-zero-vector

Table S7: Concepts excluded by the constant-coverage rule. A concept was retained only when valid in all 81 nondegenerate languages.

Statistic	Estimate	2.5%	97.5%	$B$
agreement	0.894	0.850	0.932	2000
cohen_kappa	0.856	0.794	0.908	2000
macro_f1	0.885	0.835	0.927	2000

Table S8: Agreement between the primary and reference-style concept taxonomies. Bootstrap intervals use 2,000 concept-level resamples.

Domain	Primary	External
Biological Core	68	53
Nature & Motion	53	54
Socio-Cultural	60	64
Grammar & Function	26	36

Table S9: Concept counts by taxonomy and domain before the complete-case exclusion of 11 concepts.

ID	Concept	Primary	External	Basis	Confidence
1	I	Biological Core	Grammar & Function	functional_override	high
2	you	Biological Core	Grammar & Function	functional_override	high
3	he,she,it	Biological Core	Grammar & Function	functional_override	medium
4	we	Biological Core	Grammar & Function	functional_override	high
5	you	Biological Core	Grammar & Function	functional_override	medium
6	they	Biological Core	Grammar & Function	functional_override	high

Table S10: The 22 concepts assigned to different domains by the reference-style mapping.

Domain	Dyads	Spearman $\rho$	Mean absolute difference
Biological Core	3240	0.928	0.024
Nature & Motion	3240	0.956	0.014
Socio-Cultural	3240	0.934	0.025
Grammar & Function	3240	0.764	0.051

Table S11: Dyad-level agreement of RSA dissimilarities between taxonomies.

Encoder	Mean	95% CrI
mbert	0.90973	[0.90867, 0.91079]
xlmr	0.93211	[0.93085, 0.93338]

Table S12: Global RSA multiple-membership models. Global intercepts summarize all concepts and are not Biological Core domain intercepts.

Classification	Encoder	Parameter	Estimate [95% CrI]
Primary	mBERT	alpha	0.90043 [0.89813, 0.90274]
Primary	mBERT	beta_geo	-0.00155 [-0.00426, 0.00113]
Primary	mBERT	beta_asjp	0.02220 [0.01963, 0.02478]
Primary	mBERT	beta_domain[1]	0.02405 [0.02076, 0.02728]
Primary	mBERT	beta_domain[2]	-0.02623 [-0.02951, -0.02292]
Primary	mBERT	beta_domain[3]	-0.00798 [-0.01118, -0.00476]

Table S13: Complete fixed effects for both encoders. Domain indices 1, 2, and 3 denote Nature & Motion, Socio-Cultural, and Grammar & Function; Biological Core is the reference.

Classification	Encoder	Domain	Estimate [95% CrI]
Primary	mBERT	Biological Core	0.90043 [0.89813, 0.90274]
Primary	mBERT	Nature & Motion	0.92448 [0.92215, 0.92683]
Primary	mBERT	Socio-Cultural	0.87420 [0.87186, 0.87655]
Primary	mBERT	Grammar & Function	0.89245 [0.89015, 0.89473]
Primary	XLM-R-large	Biological Core	0.93150 [0.92926, 0.93373]
Primary	XLM-R-large	Nature & Motion	0.97669 [0.97447, 0.97895]

Table S14: Adjusted mean: both encoders and both classifications. Posterior means and equal-tail 95% credible intervals reconstructed draw by draw from the frozen full-support fits.

Classification	Encoder	Domain	Estimate [95% CrI]
Primary	mBERT	Biological Core	-0.00155 [-0.00426, 0.00113]
Primary	mBERT	Nature & Motion	-0.00459 [-0.00723, -0.00187]
Primary	mBERT	Socio-Cultural	0.01138 [0.00865, 0.01409]
Primary	mBERT	Grammar & Function	0.00051 [-0.00225, 0.00321]
Primary	XLM-R-large	Biological Core	0.00583 [0.00319, 0.00848]
Primary	XLM-R-large	Nature & Motion	-0.00236 [-0.00505, 0.00031]

Table S15: Geography: both encoders and both classifications. Posterior means and equal-tail 95% credible intervals reconstructed draw by draw from the frozen full-support fits.

Classification	Encoder	Domain	Estimate [95% CrI]
Primary	mBERT	Biological Core	0.02220 [0.01963, 0.02478]
Primary	mBERT	Nature & Motion	0.02204 [0.01947, 0.02463]
Primary	mBERT	Socio-Cultural	0.02342 [0.02085, 0.02603]
Primary	mBERT	Grammar & Function	0.02566 [0.02315, 0.02820]
Primary	XLM-R-large	Biological Core	0.02107 [0.01855, 0.02357]
Primary	XLM-R-large	Nature & Motion	0.01788 [0.01540, 0.02038]

Table S16: Genealogical distance: both encoders and both classifications. Posterior means and equal-tail 95% credible intervals reconstructed draw by draw from the frozen full-support fits.

Model	$N$	Max $\widehat{R}$	Min ESS	Min E-BFMI
mBERT global (primary)	3240	1.006	434	0.642
XLM-R global (primary)	3240	1.003	644	0.633
mBERT domain (primary)	12960	1.006	533	0.704
XLM-R domain (primary)	12960	1.005	951	0.648
mBERT global (external)	3240	1.006	434	0.642
XLM-R global (external)	3240	1.003	644	0.633

Table S17: Sampling diagnostics for the non-centered, sum-to-zero multiple-membership models. R-hat and ESS summarize substantive model parameters; sampler-diagnostic columns are retained separately in the machine-readable output. The pooled rows use the complete 25,920-row support.

Layer	Domain	Mean	Geo.	95% CI	ASJP	95% CI
0	Grammar & Function	0.8789	-0.0002	[-0.0042, 0.0038]	0.0285	[0.0245, 0.0324]
0	Socio-Cultural	0.8646	0.0133	[0.0099, 0.0168]	0.0303	[0.0268, 0.0338]
1	Grammar & Function	0.8891	-0.0049	[-0.0087, -0.0011]	0.0273	[0.0235, 0.0311]
1	Socio-Cultural	0.8701	0.0121	[0.0088, 0.0154]	0.0314	[0.0281, 0.0347]
2	Grammar & Function	0.8876	-0.0048	[-0.0087, -0.0009]	0.0294	[0.0255, 0.0332]
2	Socio-Cultural	0.8608	0.0135	[0.0101, 0.0169]	0.0339	[0.0305, 0.0373]

Table S18: Layer-wise RSA sensitivity analysis for Grammar & Function and Socio-Cultural. Every row uses the same 81 languages, 196 complete-case concepts, and 3,240 dyads. Slopes and 95% ordinary least-squares intervals are descriptive and are not additional confirmatory tests.

Encoder	Predictor	Estimate	95% Wald interval
mbert	Intercept	0.93557	[0.92232, 0.94881]
mbert	Geo_Z	0.00286	[0.00221, 0.00351]
mbert	ASJP_Z	0.04151	[0.04090, 0.04211]
mbert	Frequency_Z	-0.02620	[-0.03228, -0.02012]
mbert	DomainNature_Motion	-0.00856	[-0.02240, 0.00528]
mbert	DomainSocio_Cultural	-0.00775	[-0.02079, 0.00529]

Table S19: Frequency-adjusted item-level models for both encoders. Estimates are REML fixed effects; intervals are 95% Wald intervals, not Bayesian credible intervals. Each fit uses all 635,040 observations.

Index	Language
1	Afrikaans
2	Albanian
3	Arabic
4	Armenian
5	Asturian
6	Azerbaijani

Table S20: The 81 languages retained after excluding degenerate Javanese geometry.

Item	Value
Checkpoint	xlm-roberta-large
Local checkpoint directory	models/xlm-roberta-large
Architecture	XLMRobertaModel
Layers / hidden size	24 / 1024
Layer / pooling	final; mean non-special-token states within alternative; mean across comma-delimited alternatives
Input	isolated unmarked lexical forms

Table S21: XLM-R-large extraction and fixed-sample RSA provenance. The checkpoint was loaded from local server storage without network fallback.

Encoder	Taxonomy	Domain	Mean	LOO range
mBERT	Primary	Biological Core	0.900	[0.898, 0.904]
XLM-R-large	Primary	Biological Core	0.932	[0.930, 0.933]
mBERT	Primary	Grammar & Function	0.892	[0.889, 0.895]
XLM-R-large	Primary	Grammar & Function	0.967	[0.966, 0.969]
mBERT	Primary	Nature & Motion	0.924	[0.923, 0.927]
XLM-R-large	Primary	Nature & Motion	0.977	[0.976, 0.978]

Table S22: Fixed-sample domain means and leave-one-language-out ranges for both encoders and both taxonomies. All rows use the same 81 languages, 196 concepts, and 3,240 dyads.

Taxonomy	Domain	$\rho$	Median agree.	Mean abs. diff.
Primary	All concepts	0.569	0.685	0.046
Primary	Biological Core	0.361	0.622	0.068
Primary	Grammar & Function	0.169	0.553	0.114
Primary	Nature & Motion	0.209	0.559	0.069
Primary	Socio-Cultural	0.531	0.680	0.075
External	All concepts	0.569	0.685	0.046

Table S23: Dyad-level agreement between mBERT and XLM-R-large RSA dissimilarities. Median agreement is the fraction of dyads assigned to the same side of each model’s median.

Encoder	Taxonomy	Observed domain order	Shared extremes	Full order
XLM-R-large	Primary	S-Cultural $<$ Bio. Core $<$ Grammar $<$ Nature	81/81	81/81
XLM-R-large	External	S-Cultural $<$ Bio. Core $<$ Grammar $<$ Nature	81/81	81/81
mBERT	Primary	S-Cultural $<$ Grammar $<$ Bio. Core $<$ Nature	81/81	81/81
mBERT	External	S-Cultural $<$ Grammar $<$ Bio. Core $<$ Nature	81/81	81/81

Table S24: Leave-one-language-out stability and model-specific domain ordering. Each model’s observed ordering (shown with abbreviated domain labels) was preserved in all 81 leave-one-out subsamples under both taxonomies.

Item	Value
Encoder	mBERT and XLM-R-large; separate primary/external fits plus pooled model
Python	3.12.4
NumPy / pandas	1.26.4 / 2.3.3
PyTorch / Transformers	2.4.0+cu121 / 5.14.1
CmdStan	2.38.0; invoked by the remote Python driver
Chains	4

Table S25: Software and provenance metadata for the analysis. Heavy numerical analyses were executed on the remote server; no local Bayesian sampling or model inference was used.

ID	Concept	Primary	External	External field	Confidence
1	I	Biological Core	Grammar & Function	Pronouns	high
2	you	Biological Core	Grammar & Function	Pronouns	high
3	he,she,it	Biological Core	Grammar & Function	Pronouns	medium
4	we	Biological Core	Grammar & Function	Pronouns	high
5	you	Biological Core	Grammar & Function	Pronouns	medium
6	they	Biological Core	Grammar & Function	Pronouns	high

Table S26: Complete mapping of the 207 Wiktionary-numbered study concepts. The external mapping uses a study-specific alignment to Bowern/Concepticon entries, linked WOLD/IDS fields, WordNet 3.0 lexnames where applicable, and explicit functional rules. The auxiliary alignment does not imply identity of the original concept lists; the frozen WOLD/IDS-informed macro-domain rule is retained separately from official source fields.

Mapping	Prior branch	Estimand	Mean [95% CrI]
primary	Default (0.2)	Nature & Motion - Socio-Cultural \| mBERT	0.05031 [0.04702, 0.05366]
primary	Default (0.2)	Nature & Motion - Socio-Cultural \| XLM-R-large	0.05472 [0.05128, 0.05803]
primary	Default (0.2)	(XLM-R-large - mBERT) endpoint contrast	0.00441 [-0.00034, 0.00897]
primary	Sensitivity (0.5)	Nature & Motion - Socio-Cultural \| mBERT	0.05030 [0.04695, 0.05369]
primary	Sensitivity (0.5)	Nature & Motion - Socio-Cultural \| XLM-R-large	0.05469 [0.05129, 0.05805]
primary	Sensitivity (0.5)	(XLM-R-large - mBERT) endpoint contrast	0.00439 [-0.00046, 0.00915]

Table S27: Complete pooled two-domain contrasts under default and sensitivity priors. Within-encoder and between-encoder estimands are listed separately.

Figure S1: mBERT concept-level heterogeneity. Panel (a) reports item-level regression coefficients for individual concepts. Panel (b) summarizes the full-inventory mBERT representational profile; arrows indicate observed associations. The geographic contrast shown is specific to the mBERT primary-taxonomy configuration.

Semantic Correspondence

Semantic correspondence matches equivalent points or parts across images and shapes despite changes in appearance, pose, or viewpoint.

See all 24 papers

Convergence in Science, Divergence in Religion: Calibrated Framing Differences Across Wikipedia's Language Editions
Aug 22, 2026Hung-Hsuan ChenConcept Alignment Cross-Lingual Representation Alignment
Computer Science and Information Engineering National Central University Taoyuan, Taiwan
Universal Conceptual Structure in Neural Translation: Probing NLLB-200's Multilingual Geometry
Feb 27, 2026Kyle Elliott MathewsonMultilingual Language Models Representation Geometry in Language Models
University of Alberta
Multilingual Brain Decoding Reveals Shared Cortical Semantic Geometry
Oct 17, 2026Yi Guo, Yong Hu, Michael Kwok-Po Ng +1Cross-Lingual Representation Alignment Neural Representation Geometry
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China. · University of Chinese Academy of Sciences, Beijing, China. · The University of Hong Kong, Hong Kong, China. +1

Aug 22, 2026J/K move · Enter open · S save

Convergence in Science, Divergence in Religion: Calibrated Framing Differences Across Wikipedia's Language Editions

Hung-Hsuan Chen

Computer Science and Information Engineering National Central University Taoyuan, Taiwan

Concept Alignment Cross-Lingual Representation Alignment Semantic Textual Similarity

Open paper AskPDF arXiv

When Wikipedia's language editions describe the same concept, how differently do they frame it? Prior work measures coverage gaps between editions; we measure framing distance for matched concepts. We analyze 2,799 valid articles from 3,000 possible concept-language observations, spanning 150 Wikidata-anchored concepts, 20 language editions, 4 domains, and a calibration set. Raw embedding distances reflect both content differences and how well the encoder aligns each language pair. Even among calibration concepts with stable cross-cultural denotations (e.g., chemical elements, numbers, colors), the largest language-pair mean distance is 3.6 times the smallest, and distances are typically smaller within language families. We define a baseline-adjusted distance (calibrated distance): the distance between two language versions of a concept minus the mean distance for calibration concepts in the same language pair. This adjustment substantially reduces pair-specific alignment differences and the language-family pattern. Across three multilingual encoders (LaBSE, multilingual MPNet, and CMLM), scientific articles align more closely than calibration articles, and all three rank religion first and science/technology last. Concept-level rankings are highly consistent across encoders (Spearman rho=0.75-0.79 for MPNet and CMLM relative to LaBSE). Religion lies significantly above the calibration baseline under LaBSE. Within politics, divergence concentrates on concepts such as censorship and refugee, while democracy and human rights are among the most aligned. Code, data, and per-language-pair calibration baselines are released.\footnote{https://github.com/hhchen1105/cross-linqual-concept}

Abstract

Semantic Correspondence

Explore similar work

Convergence in Science, Divergence in Religion: Calibrated Framing Differences Across Wikipedia's Language Editions

Universal Conceptual Structure in Neural Translation: Probing NLLB-200's Multilingual Geometry

Multilingual Brain Decoding Reveals Shared Cortical Semantic Geometry