Abstract

I assess the extent to which the recently introduced BERT model capturesEnglish syntactic phenomena, using (1) naturally-occurring subject-verbagreement stimuli; (2) "coloreless green ideas" subject-verb agreement stimuli,in which content words in natural sentences are randomly replaced with wordssharing the same part-of-speech and inflection; and (3) manually craftedstimuli for subject-verb agreement and reflexive anaphora phenomena. The BERTmodel performs remarkably well on all cases.

 

Quick Read (beta)