Abstracting Web Agent Proofs into Human-Level Justifications

All Papers

Information supporting answer explanations are derived from proofs. One of the difficulties in automating the explanation generation process is that the input proofs are typically described at the machine-level while explanations are often intended for human consumption. In this paper, we introduce a novel and generic approach for abstracting machine-level portable proofs into human-level justifications. This abstraction facilitates generating explanations from proofs on the web. Our approach consists of creating a repository of proof templates, called abstraction patterns, describing how machine-level inference rules and axioms in proofs can be replaced by meaningful simpler rules. Intermediate results in machine-level proofs may also be dropped during the abstraction process. The Inference Web Abstractor algorithm has been developed with the goal of matching the abstraction patterns in the repository against the original proof and applying a set of strategies to abstract the proof thereby simplifying its presentation. The tools used for creating and applying abstraction patterns are shown along with an intelligence analysis example.