Penguins Don’t Fly: Reasoning about Generics through Instantiations and Exceptions

Anonymous

Abstract: Generic statements (e.g., Birds can fly) express generalizations about the world. However, generics are not universally true -- while sparrows and penguins are both birds, penguins can't fly. Understanding cases when a generic statement is true or false is crucial for machine reasoning. In this work, we present a novel framework to generate pragmatically relevant true and false instances of a generic.We use pre-trained language models, constraining the generation based on our computational framework, and produce ${\sim}20k$ \textsc{exemplars} for ${\sim}650$ generics. Our system outperforms few-shot generation from GPT-3 (by 12.5 precision points) and our analysis highlights the importance of constrained decoding for this task and the implications of generics \textsc{exemplars} for non-monotonic reasoning and NLI.

Paper Type: long