Speakers
Description
This paper investigates whether large language models can capture the embodied dimension of word meaning by comparing LLM-generated affordance norms with human-produced norms in Czech and English. Affordance norms are systematically collected data on the actions speakers associate with concrete objects. Human data were collected from 30 Czech native speakers in a free-production experiment using 50 nouns. The same stimuli were presented to four LLMs (Claude Haiku 4.5, Gemini 3 Flash, GPT-5, and Llama 4 Maverick) in both languages, in Czech at two temperature settings; English outputs were additionally compared against human norms from Maxwell et al. (2024). We report four main findings. First, human affordance norms are more variable in Czech. Second, LLMs consistently produce narrower affordance sets than humans, with lower lexical diversity across both languages. Third, lexical overlap between human and LLM affordances is low (shared vocabulary 15–19%; Jaccard index below 0.3 at N = 10) and does not improve substantially with higher temperature or greater language representation in the training data. Fourth, human-only affordances are more embodied and emotionally grounded, while LLM-only affordances cluster around maintenance and support. The findings suggest that human norming data may offer a valuable resource for enriching lexicographic entries.