Learning Components for A Question-Answering

Abstract We describe a machine learning approach to the development of several key components in a question answering system and the way they were used in the UIUC QA system. A unified learning approach is used to develop a part-of-speech tagger, a shallow parser, a named entity recognizer and a module for identifying a question's target. These components are used in analyzing questions, as well as in the analysis of selected passages that may contain the sought after answer. The performance of the learned modules seems to be very high, (e.g., mid 90 % for identifying noun phrases in sentences), though evaluating those on a large number of passages proved to be time consuming. Other components of the system, a passage retrieval module and an answer selection module, were put together in an ad-hoc fashion and significantly affected the overall performance. We ran the system only over about 60 % of questions, answering a third of them correctly. 1 Introduction The QA system described in this paper is developed as a platform for studying and experimenting with a unified method for learning, knowledge representation and inference, required to perform knowledge intensive natural language based inferences. Our working assumption is that a robust and accurate question answering system will depend on a large number of predictors. These will be used at many levels of the process and will support a variety of functions, from knowledge acquisition to decision making and integration of information sources. Along with these, there needs to be a knowledge representation support that allows, for example, to keep track of predictions as input to higher level predictors and maintain a coherent representation of a question or a story; and, there needs to be an ability to use the outcomes of lower level predictions to make inferences that use of several of these predictors along with some constraints, e.g., those that are implied by the questions