Using similarity metrics to determine content for explanation generation

摘要

Selecting the appropriate information to convey about an entity, or collections thereof, in a knowledge or data base is a fundamental problem in explanation generation. This article reports on the development and application of a general purpose similarity metric to the task of determining content during the automatic generation of explanations. In contrast to explaining expert system reasoning (e.g., explaining how a solution was derived, justifying a solution or knowledge), we focus here on description of terms and concepts that might occur in explanations. In particular, we illustrate the application of a similarity metric to help (1) define terms and concepts in natural language text, (2) compare entities in both text and combined text and graphics, and (3) provide limited analogies in the context of a dialogue. While our algorithms only address a small portion of the content selection problem, we claim that the descriptions our algorithms generate are both precise, containing only the most significant information about a particular entity in an underlying knowledge base, and comprehensive, containing all the significant information about the entity.