Turning relational structure into a comparable number
Each concept's lexical form was encoded with two multilingual models, mBERT and XLM-R-large, averaging final-layer states over attended, non-special lexical subwords. Within each language and each concept set, vectors were centered on their language-specific mean before a cosine-dissimilarity matrix was built. Centering matters: it removes a language's overall offset so that the comparison concerns the ordering of relations rather than absolute positions in vector space.
Cross-language correspondence was then measured with representational similarity analysis. For a pair of languages, the upper-triangle entries of their dissimilarity matrices were matched in concept order and compared with a Spearman correlation; the reported index is one minus that correlation. Lower values mean more similarly ordered relations, the index runs from zero to two, and it requires no alignment of the two models' coordinate systems. Each language pair contributed one global outcome and four domain outcomes per encoder.
The four domains are author-defined operational groupings, not independently established conceptual modules: Biological Core (67 concepts), Nature & Motion (49), Socio-Cultural (59), and Grammar & Function (21). A second, reference-style grouping reassigned 22 of the 207 input concepts — for example moving personal pronouns into Grammar & Function — and the two groupings agreed on 185 input concepts, giving a 175-concept consensus subset for a robustness check.
Sociocultural relations align more closely than nature and motion
The central result is a domain contrast. Adjusted mean dissimilarity was lower for Socio-Cultural concepts than for Nature & Motion concepts in both encoders under the primary classification: 0.874 versus 0.924 for mBERT, and the same ordering held for XLM-R-large. Lower dissimilarity means the ordering of relations among those concepts corresponded more closely across languages.
The gap is reported directly as a pooled two-domain contrast. Under the primary mapping with default priors, Nature & Motion minus Socio-Cultural was 0.05031 [0.04702, 0.05366] for mBERT and 0.05472 [0.05128, 0.05803] for XLM-R-large. Under the reference-style mapping the same contrast was 0.03749 [0.03454, 0.04052] and 0.03461 [0.03163, 0.03758]. The direction and approximate size survived the regrouping, the consensus subset, label permutation, and wider prior scales.
The between-encoder comparison is a separate question and a weaker one. The endpoint contrast between XLM-R-large and mBERT was 0.00441 [-0.00034, 0.00897] under the primary mapping, an interval that includes zero. The paper's claim is therefore about a contrast that recurs across models, not about one model being demonstrably more sensitive than the other.
Geographic separation tracks divergence within sociocultural meanings
The domain contrast describes average correspondence across the whole sample. A second, distinct result concerns how that correspondence varies with distance between language communities. Two predictors were entered jointly and standardized: great-circle geographic distance between archived language centroids, and a genealogical distance from the Automated Similarity Judgment Program based on lexical–phonological dissimilarities in core vocabulary. Because both are in the same model, each slope describes association per standard-deviation increase in one distance, conditional on the other.
Within the Socio-Cultural domain, relational differences increased with geographic separation in both encoders. For mBERT the Socio-Cultural geographic slope was 0.01293 [0.00940, 0.01643]; for XLM-R-large it was 0.00226 [-0.00113, 0.00562], positive in mean but with an interval touching zero. Less closely related languages also differed more across all domains, consistent with the positive genealogical slopes. The authors are explicit that geographic separation does not directly measure contact, and that word-form dissimilarity does not isolate inheritance — the proxy can reflect both inheritance and borrowing, and it shares material with the semantic inventory.
These two findings answer different questions and should not be merged. A domain can be relatively similar on average across the sample and still show systematic regional differentiation. The paper's interpretation is that broad recurring relational patterns coexist with regional conventions, but the distance associations are correlational and the mechanisms — communicative need, contact, ecology, history — were not independently measured.
What the contrast does not establish about speakers
The evidence is model-derived lexical relations. Word embeddings represent regularities in linguistic context as relations in a quantitative space, and the paper justifies their use by pointing to evidence that distributional structure captures aspects of human semantic similarity. But the study collected no behavioral or neural responses, recruited no participants, and did not test whether speakers of these languages actually judge the concepts as more alike in the sociocultural domain. The authors state that establishing correspondence to speakers' judgments remains necessary, particularly because naming and nonverbal similarity can dissociate.
Several further boundaries sit close to the results. The domains are operational groupings rather than validated modules, and the regrouping exercise shows that assignment choices move individual concepts between domains even when the headline contrast survives. The language sample follows usable common coverage, so it is not representative of the world's languages. Training biases, polysemy, and model-dependent intermediate ranks affect generalization. The models also omitted an additional dyad effect and residual correlations across outcomes from the same dyad, so the reported intervals are conditional on that specification.
The paper's own forward-looking claim is a prediction rather than a finding: matched human semantic judgments should show greater cross-language correspondence for the sampled Socio-Cultural relations than for Nature & Motion relations. Comparisons with nonverbal materials and manipulations of language use are named as the tests that would connect this prediction to attention, categorization, and learning.
Where the relativity question should be tested next
The paper's contribution to the linguistic relativity debate is a comparative profile rather than a verdict. It argues that universality and relativity are interwoven across conceptual domains: shared basic meanings can participate in relations whose cross-language correspondence varies with content. On this account, shared cognitive resources interact with linguistic and local experience, and the interesting question shifts from whether languages differ to where within the lexicon their relational structures converge and diverge.
For readers weighing this evidence, the priority is the domain split, not a single global similarity score. The sociocultural-versus-nature-and-motion contrast is the result that recurred across two encoders, two classifications, and the consensus subset, and it is the one the authors turn into a testable prediction about human judgments. The geographic gradient within sociocultural meanings is a secondary, correlational pattern whose encoder-specific intervals differ in strength. The next decisive evidence would be matched speaker judgments on the same concept sets, because that is the step this design deliberately leaves open.
Sources used in this answer
Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages
Across 81 languages and 196 basic concepts, relations among sociocultural meanings corresponded more closely across languages than relations among nature and motion meanings, a contrast that recurred across two multilingual word-embedding models and two concept classifications.