Knowledge base refinement by backpropagation

摘要

This paper presents a novel approach to knowledge base refinement based upon a neural-network computational model known as backpropagation. In this approach, the rule base of a knowledge-based system is converted to a neural network by performing a fine-grained mapping: each domain attribute or concept is mapped into a neural node and each rule is mapped into a connection. Neural networks thus constructed are more sparse in connections that conventional ones. This nature in conjunction with sufficient initial knowledge embedded in such networks accounts for their faster convergence to a desired state in the learning phase. The direction along with the magnitude of changes in connection weights after training the networks with correct samples is the main basis to revise corresponding rule bases. The approach is further strengthened by employing overall system performance measures such as diagnostic accuracy to approve or disapprove modifications made by backpropagation.