Dynamic Neuro-Symbolic Knowledge Graph Construction for Zero-shot Commonsense Question Answering

Understanding narratives requires reasoning about implicit world knowledge related to the causes, effects, and states of situations described in text. At core this challenge is how access contextually relevant on demand reason over it. In paper, we present initial studies toward zero-shot commonsense question answering by formulating task as inference dynamically generated graphs. contrast previous for integration that rely retrieval existing from static graphs, our study where often not bases. Therefore, a novel approach generates contextually-relevant symbolic structures using generative neural models. Empirical results two datasets demonstrate efficacy neuro-symbolic constructing graphs reasoning. Our achieves significant performance boosts pretrained language models vanilla models, all while providing interpretable paths its predictions.