What Don't RNN Language Models Learn About Filler-Gap Dependencies?
Rui Pedro Chaves · ScholarWorks@UMassAmherst (University of Massachusetts Amherst) · 2020
In a series of experiments, Wilcox et al. (2019,2019) provide evidence suggesting that general-purpose state-of-the-art LSTM RNN language models have not only learned English filler-gap dependencies, but also some of their associated 'island' constraints (Ross 1967). In the present paper, I cast doubt on such claims, and argue that upon closer inspection filler-gap dependencies are learned only very imperfectly, including their associated island constraints. I conjecture that the LSTM RNN models in question have more likely learned some surface statistical regularities in the dataset rather than higher-level abstract generalizations about the linguistic mechanisms underlying filler-gap constructions.