Learning to ask relevant questions

https://doi.org/10.1016/S0004-3702(99)00037-5Get rights and content
Under an Elsevier user license
open archive

Abstract

This paper describes an effective technique for relevant questioning in expert systems whose knowledge base is encoded in a propositional formula in conjunctive normal form. The methodology does not require initial knowledge about the relationships between questions. Instead, the system learns such relationships over time as follows. After each session, the system analyzes its questioning, deduces how it could have obtained each conclusion without asking irrelevant questions, and records the relevant questions and answers in so-called processed dialogues. When a question is to be selected in a subsequent session, the system measures the relevancy of questions using the processed dialogues, ranks the questions according to that measure, and asks the highest-ranked question next. We have used the methodology in an expert system that handles industrial chemical exposure management. In that application, the system learned rather quickly to ask relevant questions and became just as effective as a human expert.

Keywords

Intelligent interfaces
Algorithms
Learning interaction models
Knowledge representation
Algorithms
Logic programming and theorem proving

Cited by (0)

This research was supported in part by ONR grant N00014-93-1-0096 and AASERT DOD grant N00014-95-1-0990.