Actions, answers, and uncertainty: a decision-making perspective on Web-based question answering

We present research on methods for generating answers to freely posed questions, based upon information drawn from the Web. The methods exploit the typical redundancy of information on the Web by making multiple queries to search engines and then combining the search results into an answer. We focus on the pursuit of techniques for guiding information gathering in support of answering questions via the learning of probabilistic models that predict the value of information drawn from the Web. We first review research on question-answering systems. Then, we present AskMSR, a prototype Web-based question-answering system. We describe the learning of Bayesian-network models that predict the likelihood that answers are correct, based on multiple observations. We review a two-phased Bayesian analysis and present an expected-utility analysis of information-gathering policies using these inferences. After reviewing the results of a set of experiments, we describe research directions.