Bridging Knowledge Gaps in Neural Entailment via Symbolic Models