Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages
Do people who speak different languages organize concepts in similar ways? The authors extract word embeddings from multilingual transformers (mBERT and XLM-R-large) to compare how 196 basic concepts relate to one another across 81 languages. Their core finding is that sociocultural meanings are more aligned across languages than nature and motion meanings, a domain-sensitive pattern that recurs across encoders and concept groupings. The work connects computational linguistics to cognitive science, though its inferences about human conceptual organization remain indirect because they rely entirely on neural network representations rather than behavioral or neural data from speakers.
The paper offers a carefully executed, domain-sensitive computational exploration of linguistic relativity, but its conclusions about human cognition are limited by its exclusive reliance on pretrained multilingual embeddings without independent human validation. The authors are commendably transparent about this limitation, noting that model-derived lexical relations still need to be connected to speakers' judgments. The Socio-Cultural versus Nature & Motion contrast is robust across two encoders and multiple robustness checks, yet the inferential leap from vector geometry to conceptual organization remains unsupported by direct evidence.
The core domain contrast is the paper's most defensible result: relations among sampled Socio-Cultural concepts correspond more closely across languages than those among Nature & Motion concepts, and this ordering recurs in both mBERT and XLM-R-large, survives regrouping, and holds under permutation testing. The statistical approach is also principled, with Bayesian regressions that properly include language random effects to handle dyadic dependence, and the authors document quality-control exclusions honestly.
The central inferential gap is that pretrained multilingual embeddings are treated as proxies for human conceptual structure without independent validation. The authors acknowledge this but do not close it. A further limitation is that the four domains are admitted to be 'operational groupings, rather than independently established conceptual modules,' which weakens any strong theoretical claims about distinct cognitive modules. There is also a circularity risk: the ASJP genealogical distance measure shares core-vocabulary material with the semantic inventory, potentially inflating the association between word-form relatedness and semantic alignment. The authors additionally note that 'Models omitted an additional dyad effect and residual correlations across outcomes from the same dyad; intervals are conditional on this specification,' meaning the reported uncertainties may be optimistically narrow.
The comparison to prior work is fair; the authors correctly relate their finding to Thompson et al. (2020) and Lewis et al. (2023). However, the two-encoder design is only partially independent reassurance, because mBERT and XLM-R-large share a multilingual masked-language-modeling objective and overlapping training corpora, which could artificially inflate cross-lingual correspondence. The pooled between-encoder difference in the key Nature & Motion minus Socio-Cultural contrast is small and its 95% credible interval includes zero, so agreement concerns only the direction, not the magnitude. The paper also fairly cautions that 'Extraction choices matter for interpreting these representations,' reminding readers that the exact pooling and centering rules influence the numerical outcomes.
The authors share data, scripts, and analysis specifications via Figshare, which supports transparency. However, several factors would complicate exact independent reproduction. The archived model checksums are incomplete: the authors state that 'Checkpoint weights are not included in the manuscript package' and that 'Complete weight-hash coverage is therefore not claimed.' In addition, the Wiktionary source material is dynamic, and while the authors retained archived lexical forms, reproduction requires using those exact archived snapshots rather than current downloads. The quality-control pipeline also involves language-specific exclusion rules—such as the Javanese duplicate-threshold cut and zero-vector handling—that are documented but would need to be reimplemented precisely.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.