Counterfactual reasoning: Testing language models' understanding of hypothetical scenarios
This work addresses the challenge of evaluating language models' understanding of hypothetical scenarios for AI researchers, though it is incremental in probing existing models.
The study tested five pre-trained language models on counterfactual reasoning tasks to distinguish statistical correlation from systematic logical reasoning, finding that models could override real-world knowledge in hypothetical scenarios, with GPT-3 showing some sensitivity to linguistic nuances.
Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on the understanding of real world. We tease these factors apart by leveraging counterfactual conditionals, which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from five pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge -- however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.