Differentiable reasoning over symbolic KB

We describe two novel methods, NQL and EmQL, of representing symbolic knowledge bases (KB) that enables neural KB inference. The two methods are fully differentiable, faithful to the original semantics of the KB, expressive enough to model multi-hop inferences, scalable enough to use with realistically large KBs, and potentially capable of generalizing to novel KB entries. We show experiment results on several KB based tasks including KB completions and KB question answering (KBQA). Besides KB-based tasks, we show that the two KB reasoning modules can be integrated into language models to perform challenging open-domain QA tasks.