In Proceedings of the International Conference on Logic Programming , 1999

Existing explanation systems for deductive databases show forests of proof trees. Although proof trees are often useful, they are only one possible interesting representation. We argue that an explanation system for deductive databases must be able to generate explanations at several levels of abstraction. One possible and well known technique to achieve this exibility is to instrument meta-interpreters. It is, however, not often used because of its ineeciency. On the other hand, deductive databases often generate intermediate information stored in the physical database. This information can be considered as a low-level trace giving a faithful picture of what has happened at the relational level. The deductive reasoning is lost but can be very easily recovered by a meta-interpreter. In this article we describe a technique to generate explanations by integrating a relational trace and an instru-mented meta-interpreter. The expensive aspects of meta-interpretation are reduced by the use of the trace which avoids many costly calculations. The exibility of meta-interpretation is preserved, as illustrated by the generation of three diierent kinds of explanations: a box-oriented trace, a multi-SLD-AL tree and abstract AND trees. This technique enables powerful explanation systems to be implemented with very few modiications of the deductive database mechanism itself.