From Parallels to Divergence: Language Models' Learning of Action Nominal Constructions
Abstract
This paper investigates the impact of grammatical configurations on language models' learning of action nominal constructions (ANCs) and compares these learning preferences with typological distributions in natural languages.
Reality Card
The study reveals that while language models show typological parallels in basic clause configurations, they exhibit divergences in complement systems and ANC strategies, learning combinations that are rare or unattested in natural languages.
The authors constructed 96 target languages and evaluated models using targeted minimal pairs covering 34 grammatical phenomena.
The grammar-engineered method may limit reproducibility due to its reliance on synthetic corpus construction.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.