Look at that! BERT can be easily distracted from paying attention to morphosyntax


Syntactic knowledge involves not only the ability to combine words and phrases, but also the capacity to relate different and yet truth-preserving structural variations (e.g. passivization, inversion, topicalization, extraposition, clefting, etc.), as well as the ability to infer that these syntactic variations all adhere to common morphosyntactic rules, like subject-verb agreement. Although there is some evidence that BERT has rich syntactic knowledge, our adversarial approach suggests that it is not deployed in a robust and linguistically appropriate way. English BERT can be tricked to miss even quite simple syntactic generalizations, when compared with GPT-2, underscoring the need for stronger priors and for linguistically controlled experiments in evaluation.

Proceedings of the Society for Computation in Linguistics