CLASP
The Centre for Linguistic Theory and Studies in Probability

Discontinuous Constituency and BERT: Case Studies of Dutch

Abstract

Large-scale language models like GPT-3 and BERT are considered stochastic parrots: they make reference to form but not to meaning. Yet, such models allow for high-performing fine-tuned models in NLP, begging the question of the how much linguistic information we can extract from them. In this talk, I discuss recent case studies in which we set out to quantify the syntactic capacity of BERT in the evaluation regime of non-context free patterns, as occurring in Dutch. We devised test suites based on a mildly context-sensitive formalism, from which we derive grammars that capture the linguistic phenomena of control verb nesting and verb raising, and other verb cluster formations. The grammars, paired with a small lexicon, provide us with a large collection of naturalistic utterances annotated with verb-subject pairings, that serve as the evaluation test bed for an attention-based span selection probe. Our results, backed by extensive analysis, suggest that the models investigated fail in the implicit acquisition of the dependencies examined. If time permits, we compare the two different strategies for generating test cases, and look into future extensions of this work.