Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Do Concreteness Norms Transfer Across Languages? Evidence from Hebrew, English, and Large Language Models
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Tammar Truzman1, Netaniel Rein1, Ajay Halai1; 1Cognition and Brain Sciences Unit, University of Cambridge
*Background*: Concreteness, defined as the extent to which a word’s meaning is grounded in perception and action (Muraki et al., 2022), is a core psycholinguistic property widely used in cognitive neuroscience, psycholinguistics, and clinical language research. Large-scale concreteness norms exist primarily for English(Muraki & Pexman, 2025), whereas resources in other languages remain limited. In Hebrew, norms are currently restricted to a small subset of nouns (Stein et al., 2024), limiting control of psycholinguistic variables in research and clinical assessment. Although translated English norms are often used, the extent to which concreteness transfers across languages and lexical categories remains unclear. Recent work further suggests that large language models (LLMs) may provide scalable estimates of concreteness, though their reliability across languages and parts of speech remains unknown. Here, we hypothesised that (i) concreteness would transfer across languages, but more strongly for nouns than verbs, and (ii) LLM-derived estimates would show human-like cross-linguistic agreement and reproduce this dissociation/// *Methods*: Thirty-six native Hebrew speakers and 36 native English speakers provided concreteness ratings for 200 translation-matched infinitive verbs(Masterson & Druks, 2015). The Hebrew group additionally rated 60 nouns to enable comparison with existing Hebrew norms (Drori, n.d.). All instructions were adapted from Brysbaert et al.(2013). GPT4o-derived concreteness values were generated for the same word sets in both Hebrew and English across multiple runs and prompting conditions, following the methodology of Martínez et al.(2024). Internal reliability was assessed using random split-half analyses and bootstrap resampling procedures. Associations across lexical categories, languages, and human–LLM ratings were examined using correlations, with bootstrap- and Fisher z-based inferential comparisons where appropriate. Follow-up analyses examined the effects of translation, word ambiguity, and verb morphology on observed cross-language differences/// *Results*: Hebrew noun ratings showed very high reliability across raters (split-half ρ ≈ .93) and strong agreement with both existing Hebrew norms (Drori; ρ ≈ .93) and English concreteness norms (Brysbaert; ρ ≈ .90). In contrast, verbs showed lower reliability and greater variability within-language (Hebrew: ρ ≈ .79; English: ρ ≈ .84). Cross-language agreement for verbs was substantially reduced (Hebrew–English: ρ ≈ .75–.77), and correlations with Brysbaert norms varied across morphological verb forms (e.g., base, -ing, past tense; ρ ≈ .67–.84). Restricting analyses to low-ambiguity verbs did not improve correlations. As expected, GPT-derived values demonstrated high internal consistency across runs (ρ ≈ .96–.99) and moderate-to-strong agreement with human ratings. Critically, GPT correlations were similar to those reported by Martínez et al. (2024) and reproduced the same noun–verb dissociation observed in human data, with substantially stronger performance for nouns (ρ ≈ .84–.89) than verbs (ρ ≈ .67–.77)/// *Conclusions*: Concreteness transferred robustly across languages for nouns, but substantially less so for verbs, suggesting important lexical-category differences in semantic representation and stability. The findings further suggest that translated norms and LLM-derived estimates may provide useful approximations for noun concreteness in under-resourced languages, supporting the development of better-controlled experimental and clinical language assessments. In contrast, verbs and other parts of speech require direct empirical validation. All ratings, code, and resources will be made openly available via OSF.
Topic Areas: Computational Approaches, Meaning: Lexical Semantics