An Intelligent System for Question Answering

Sanda M. Harabagiu Dan I. Moldovan University of Southern California Southern Methodist University Los Angeles, CA 90089-2562 Dallas, Texas 75275-0122 harabagi@usc.edu moldovan@seas.smu.edu Abstract This paper introduces a system intended for question answering based on abductive inference. The system uses a large knowledge base structured around WordNet. Texts and questions are uniformly processed and semantic paths between concepts are established using a marker propagation method. The paths bring forward inferences and implicatures that are otherwise di cult to extract. Implicit inferences contribute to the information provided by explicit inferences, producing answers for a variety of question types. 1 Basic Idea An approach for building intelligent systems is to implement an explicit declarative knowledge base on which operates a formal reasoning mechanism. Thus, of central importance are: (1) knowledge representation and structuring, (2) design of inference rules, and (3) control of inference process. Since the entire system revolves around the knowledge base, for many, intelligent systems are synonymous with knowledgebased systems. Experience has shown that there are major problems with this approach; only modest size systems working in restricted domains were successfully implemented. Since knowledge is unlimited, no matter how large the knowledge base is, it is not possible to capture all the knowledge in the world. Knowledge based systems may work well in some domain, but quickly fail when the domain changes slightly. Lenat [6] called this the brittleness problem. Text-based Intelligent Systems Another possible solution for intelligent systems, that becomes more and more evident and feasible, is to use the expressive power of natural language to represent knowledge. Jacobs [5] called this text-based intelligent systems ( TBIS ) and was among the rst to suggest that this approach may replace or augment traditional knowledge-based systems. In essence, a TBIS may work like this: a text that represents domain, context or any other form of information is transformed by a natural language processing system into a knowledge representation. The sophistication of natural language processing may vary from simple tagging to \deep understanding", and this determines the power of such a system. Then, the user provides information also in the form of natural language. This information may be a query, or some other form. The user's information needs to be transformed by the same natural language processing system in a form of knowledge. Then, the text and user representations are compared, and inferential information is extracted.