Unbounded knowledge acquisition based upon mutual information in dependent questions

This paper describes an experimental system for knowledge acquisition based on a general framework exemplified in the game of twenty questions. A sequence of propositional questions is put to the user in an attempt to uncover some hidden concept, and the answers are used to expand and refine the system’s knowledge of the world. Previous systems adopting this framework typically represent knowledge as a matrix of truth values or weights that relate entities to attributes—such that if the hidden concept is “a bird”, for example, then the answer to a question about whether the target entity can fly is based on the extent to which “flying” is generally attributable to “a bird” as measured by the value in the matrix element indexed by the attribute-entity pair. Our system adopts a subtly different approach wherein knowledge is a measure of the extent to which answers to pairs of questions are co-dependent. Thus, knowledge about birds being able to fly is captured by the mutual information in the answers to a pair of questions like “Can it fly?” and “Is it a bird?”. We present a case that this offers a practical and epistemologically sound basis for acquiring knowledge.
Conference Contribution
Type of thesis
Smith, T.C. & van de Molen, C. (2011). Unbounded knowledge acquisition based upon mutual information in dependent questions. In J. Li (Ed.): AI 2010: Advances in Artificial Intelligence, Proceedings of the 23rd Australasian Joint Conference, Adelaide, Australia, December 7-10, 2010 (pp. 233-242). Springer-Verlag Berlin Heidelberg.