Abductive Explanations for Text Understanding: Some Problems and Solutions. Technical Report Number of Training Examples Flex-either Id3 7.2 Multiple Categories: Soybean Results

A Prolog-like inference system for computing minimum-cost abductive explanations in natural-language interpretation. 27 [Michalski and Chilausky, 1980] R. S. Michalski and S. Chilausky. Learning by being told and learning from examples: An experimental comparison of the two methods of knowledge acquisition in the context of developing an expert system for soybean disease diagnosis. Acknowledgements We would like to thank Mick Noordewier and Jude Shavlik for providing the DNA theory and data and helping us interpret the results; Je Mahoney for translating the soybean theory and data and implementing the exible matcher; and Hwee Tou Ng for providing the abduction component. 24 since generalizing or learning a rule that concludes P actually specializes the overall theory by preventing this antecedent from being satised. Conversely, specializing or eliminating a rule for P may actually generalize the overall theory. Therefore, the system will have to consider standard generalization operators as specializers in certain contexts and vice versa. Third, the current system assumes all examples are instances of exactly one of the top-level categories. It cannot directly accept examples of intermediate concepts nor deal with overlapping categories. A truly robust theory revision system should be able to accept examples of any of its concepts and use them to revise the rules for that concept directly or to revise other concepts indirectly. Fourth, Either is basically restricted to revising propositional theories. We have already developed a prototype of a successor system called Forte which is capable of revising rst-order Horn-clause theories [Richards and Mooney, 1991]. This system is undergoing continued development to improve its eciency and capabilities. Finally, Either is restricted to revising purely logical theories and many rule bases employ some form of probabilistic reasoning. Some previous work has addressed the problem of rening the probabilities or certainty factors attached to rules [Ling and Valtorta, 1991; Ginsberg et al., 1988]; however, such numerical adjustments have not been integrated with more symbolic revisions such as learning new rules. We are currently developing a system that rst \tweaks" certainty factors until no more improvement is possible and then resorts to learning new rules. The system cycles between \tweaking" and rule learning until it converges to 100% accuracy on the training data. 10 CONCLUSIONS The development and testing of Either has demonstrated how deduction, abduction, and induction can be successfully integrated to revise imperfect domain theories. By combining such diverse methods, Either is able to handle a …