Neuro-symbolic Natural Logic with Introspective Revision for Natural Language Inference

Abstract We introduce a neuro-symbolic natural logic framework based on reinforcement learning with introspective revision. The model samples and rewards specific reasoning paths through policy gradient, in which the revision algorithm modifies intermediate symbolic steps to discover reward-earning operations as well leverages external knowledge alleviate spurious training inefficiency. is supported by properly designed local relation models avoid input entangling, helps ensure interpretability of proof paths. proposed has built-in shows superior capability monotonicity inference, systematic generalization, interpretability, compared previous existing datasets.