Agenda
Michelle Suijkerbuijk – The success of Neural Language Models on syntactic island effects is not universal: strong wh-island sensitivity in English but not in Dutch
A much-debated question in linguistics is whether learning language requires a language-specific learning capacity or can arise from input alone. Neural language models (NLMs) can greatly influence this debate as they learn solely from input and their inductive biases, without built-in linguistic representations. Recent work has examined how NLM’s handle various grammatical phenomena, including syntactic island constraints, which block filler-gap dependency formation in structures like wh-phrases.
(1) *Whati do you wonder [wh whether John bought _i]?
Although island violations like (1) rarely occur in language input and NLMs cannot fall back on built-in linguistic knowledge, studies show that they can model these constraints successfully. However, these studies typically only assume a correspondence between model and human behavior, and focus almost exclusively on English. This study addresses these gaps by directly comparing NLMs’ and humans’ behavior on wh-islands in English and Dutch.
We introduce two key improvements. First, five models and >70 human participants were presented with the same sentences manipulated for presence of island, gap, and filler. By comparing model-assigned probabilities with human acceptability judgments, we test whether NLMs represent a wh-island sensitivity in a human-like way. Second, we extend this approach cross-linguistically, contrasting English (SVO) and Dutch (SOV), two related languages differing in word order. Our results replicate previous findings that NLMs show a wh-island sensitivity comparable to English participants. However, the same pattern does not generalize to Dutch: while Dutch participants showed a strong wh-island sensitivity similar to English participants and models, Dutch NLMs did not. These findings suggest that NLMs’ apparent success on English does not straightforwardly extend across languages. Cross-linguistic evidence is therefore crucial before NLMs can be claimed to bear on the human capacity for grammar learning.
Michelle Suijkerbuijk is affiliated with the Centre for Language Studies, Radboud University
Teams
Meeting ID: 364 963 289 337 5
Passcode: me7HB6rG