The authors provide an innovative approach to information retrieval (IR), based on runtime selection of the best set of techniques to respond to a given query. A computational model is developed for fast runtime selection of the best IR techniques, and a case study of building a predictive probabilistic model of the performance of an IR technique is presented. The authors also develop effective mechanisms for the runtime selection of the best IR techniques, based on a progressive processing model. The paper further discusses a method for running artificial neural networks that take query characteristics as input, and output whether a certain IR technique is expected to improve or degrade the quality of retrieved documents, if run on the query.
A new approach to representing the problem within the progressive processing framework is developed. The resulting meta-level control was solved by reformulating it as a Markov decision problem. The authors develop a fast approximation of the opportunity cost that allows a reactive controller to select the best IR techniques using a library of pre-compiled control policies. They also examine the ability to predict the performance of IR techniques using local context analysis (LCA), as a case study. The authors show that time-consuming IR techniques can be integrated in a robust way into systems. Future research work is discussed in the paper, and an extensive literature review is provided.