What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models
This work addresses the need for better evaluation tools in NLP to understand model limitations, though it is incremental as it applies existing psycholinguistic methods to a specific model.
The authors tackled the problem of understanding what linguistic capacities BERT acquires through pre-training by introducing a new suite of psycholinguistic diagnostics drawn from human experiments. They found that BERT can distinguish good from bad completions involving shared categories or role reversals (with less sensitivity than humans) and retrieves noun hypernyms robustly, but struggles with challenging inferences, role-based event prediction, and shows clear insensitivity to negation impacts.
Pre-training by language modeling has become a popular and successful approach to NLP tasks, but we have yet to understand exactly what linguistic capacities these pre-training processes confer upon models. In this paper we introduce a suite of diagnostics drawn from human language experiments, which allow us to ask targeted questions about the information used by language models for generating predictions in context. As a case study, we apply these diagnostics to the popular BERT model, finding that it can generally distinguish good from bad completions involving shared category or role reversal, albeit with less sensitivity than humans, and it robustly retrieves noun hypernyms, but it struggles with challenging inferences and role-based event prediction -- and in particular, it shows clear insensitivity to the contextual impacts of negation.