The success of Neural Language Models on syntactic island effects is not universal: wh-island sensitivity in English but not in Dutch

A much-debated question in linguistics is whether learning language, specifically grammar, requires a language-specific learning capacity or can be achieved from input alone. The current study adds to this debate by testing whether Neural language models (NLMs), which learn solely from input and their non-linguistic inductive biases, without any built-in linguistic representations, can model the wh-island constraint. We improve on previous research by testing both NLMs and human participants in the same experimental set-up and directly comparing their results, and doing this cross-linguistically in both English and Dutch. Our results show that the NLMs can successfully model the wh-island constraint in English but not in Dutch, opposing the arguments in favor of learning from input alone.