Does BERT really agree ? Fine-grained Analysis of Lexical Dependence on a Syntactic Task

Anonymous

Abstract: Although transformer-based Neural Language Models obtain impressive results on a wide variety of tasks, their generalization abilities are not well understood. They have been shown to perform strongly on subject-verb number agreement in a wide array of settings, suggesting that they learned to capture syntactic dependencies during their training even without explicit supervision. In this paper, we examine the extent to which BERT relies on lexical content to solve the number agreement (NA) task. To do so, we disrupt the lexical patterns found in naturally occurring stimuli in a novel fine-grained analysis of BERT's behavior. Our results on nonce sentences suggest that the model generalizes well for simple structures, but fails to perform lexically-independent syntactic generalization when as little as one attractor is present.