Are Rotten Apples Edible? Challenging Commonsense Inference Ability with Exceptions

Abstract

Previous studies have argued that pre-trained language models encode commonsense relational knowledge (e.g. that apples are edible). However, simultaneous work has revealed that such models are often insensitive to context, even ignoring overt contextual cues such as negations. In this paper, we investigate whether masked language models (the BERT family) can move beyond naive associative biases (e.g., apple ? edible) when the context warrants (e.g. ranking inedible higher when presented with the information that the apple is rotten). We introduce the WINOVENTI procedure, which adversarially exploits generic associations in masked language models to create model-specific Winograd-style entailment schemas. Using our constructed WINOVENTI challenges set of over 2, 000 schemas, we show that language models in the BERT family experience a steep drop in performance on prompts that require them to pick answers which require reasoning about context (e.g., from 89.8% to 18.4% for BERTLARGE). We present evidence that language models exhibit different associative biases, suggesting a need for future work in developing and analyzing frameworks similar to WINOVENTI that are tuned to model-specific weaknesses.