Knowledge-Based Question Answering

This paper describes the Webclopedia Question Answering system, in which methods to automatically learn patterns and parameterizations are combined with hand-crafted rules and concept ontologies. The source for answers is a collection of 1 million newspaper texts, distributed by NIST. In general, two kinds of knowledge are used by Webclopedia to answer questions: knowledge about language and knowledge about the world. The former is embodied both in the Information Retrieval engine that identifies likely answer sentences and the CONTEX parser that analyses the input question and candidate answers. The latter is embodied in a QA typology of some 140 nodes, a concept ontology of some 10,000 concepts, and the IR engine's ranking algorithm that takes typical document structure into account.