Leap-Of-Thought: Teaching Pre-Trained Models to Systematically Reason Over Implicit Knowledge

To what extent can a neural network systematically reason over symbolic\nfacts? Evidence suggests that large pre-trained language models (LMs) acquire\nsome reasoning capacity, but this ability is difficult to control. Recently, it\nhas been shown that Transformer-based models succeed in consistent reasoning\nover explicit symbolic facts, under a "closed-world" assumption. However, in an\nopen-domain setup, it is desirable to tap into the vast reservoir of implicit\nknowledge already encoded in the parameters of pre-trained LMs. In this work,\nwe provide a first demonstration that LMs can be trained to reliably perform\nsystematic reasoning combining both implicit, pre-trained knowledge and\nexplicit natural language statements. To do this, we describe a procedure for\nautomatically generating datasets that teach a model new reasoning skills, and\ndemonstrate that models learn to effectively perform inference which involves\nimplicit taxonomic and world knowledge, chaining and counting. Finally, we show\nthat "teaching" models to reason generalizes beyond the training distribution:\nthey successfully compose the usage of multiple reasoning skills in single\nexamples. Our work paves a path towards open-domain systems that constantly\nimprove by interacting with users who can instantly correct a model by adding\nsimple natural language statements.\n