Framework for Answering Queries Using Multiple Representation and Inference Techniques.

摘要

The Polylog framework is designed to provide a language for efficiently automating complex queries of information represented in multiple formats. A Polylog program contains a set of modules called specialists that store and make inferences about data in a particular representation. The FocusLoop algorithm answers queries by combining the knowledge and computation of all the specialists. Logic program duals for Polylog programs are introduced to prove that FocusLoop is sound and complete. A logic program dual makes the same inferences as the Polylog program it corresponds to. By using one program to formally characterize behavior and another to implement it, the traditional tradeoffs between provably correct automated question answering, representational flexibility and efficient execution are greatly reduced. Specialists using representations such as neural networks, ontologies, logical clauses and constraint graphs have already been implemented. They demonstrate that complex queries over multiple data sources can be automated without sacrificing efficiency for soundness and completeness. Finally, it is shown that FocusLoop generalizes logical deduction using operations such as resolution, forward inference and subgoaling and that these are common themes in many computational frameworks. In Polylog, each operation is implemented using multiple algorithms, enabling the weaknesses and impasses of one inference or representation technique to be compensated for by the strengths and resources the others.