Daniel J. McCloskey

Genetic Programming for Feature Selection and Question-Answer Ranking in IBM Watson

IBM Watson is an intelligent open-domain question answering system capable of finding correct answers to natural language questions in real-time. Watson uses machine learning over a large heterogeneous feature set derived from many distinct natural language processing algorithms to identify correct answers. This paper develops a Genetic Programming (GP) approach for feature selection in Watson by evolving ranking functions to order candidate answers generated in Watson. We leverage GP’s automatic feature selection mechanisms to identify Watson’s key features through the learning process. Our experiments show that GP can evolve relatively simple ranking functions that use much fewer features from the original Watson feature set to achieve comparable performances to Watson. This methodology can aid Watson implementers to better identify key components in an otherwise large and complex system for development, troubleshooting, and/or customer or domain-specific enhancements.