arXiv paper: Evolving language compositionality in a frequency-structured meaning space

Researchers modified the classic iterated learning model so that some whole meanings appeared much more frequently than others. They found that high-frequency meanings can resist grammatical regularization, mirroring irregular forms in natural languages like frequent verbs. Crucially, when frequency was instead applied to parts of meanings—such as individual features—the emerging language collapsed and failed to transmit across generations, because learners could not piece together a compositional system from fragmentary high-frequency elements alone. This suggests that frequency distributions can only shape durable linguistic structure when they are defined over whole form–meaning units that learners can acquire holistically; distributing frequency over smaller sub-units undermines the relational patterns needed for compositional generalization and stable cultural transmission of language.