Laurie Hiyakumoto, Lucian Vlad Lita, and Eric Nyberg, Carnegie Mellon University
This paper presents a flexible approach using a utility-based planner to choose between different extraction methods in a question-answering (QA) system, enabling multiple run-time strategies. We model the QA process as a set of probabilistic actions, and use the planner to select an action sequence maximizing the expected value of information. The planner exploits redundancy and diversity by dynamically selecting among three statistical and pattern-based answer extractor actions in an end-to-end QA system. The expected value of information produced by each extractor is modeled as a function of the current question context. Experiments demonstrate that a planning-based approach can outperform a fixed strategy using the single-best extractor.