T he closely related fields of question answering and information extraction aim to search large databases of textual material (textbases) to find specific information required by the user (1, 2). As opposed to information retrieval systems, which attempt to identify relevant documents that discuss the topic of the user's information need, information extraction systems return specific information such as names, dates, or amounts that the user requests. Although information retrieval systems (such as Google and Alta Vista) are now in widespread commercial use, information extraction is a much more difficult task and, with some notable exceptions, most current systems are research prototypes. However, the potential significance of reliable information extraction systems is substantial. In military, scientific, and business intelligence gathering, being able to identify specific entities and resources of relevance across documents is crucial. Furthermore, some current information extraction systems now attempt the even more difficult task of providing summaries of relevant information compiled across a document set.The majority of current information extraction systems are based on surface analysis of text applied to very large textbases. Whereas the dominant approaches in the late 1980s and early 1990s would attempt deep linguistic analysis, proposition extraction, and reasoning, most current systems look for answer patterns within the raw text and apply simple heuristics to extract relevant information (3). Such approaches have been shown to work well when information is represented redundantly in the textbase and when the type of the answer is unambiguously specified by the question and tends to be unique within a given sentence or sentence fragment. Although these conditions often hold for general knowledge questions of the kind found in the Text REtrieval Conference (TREC) Question Answer track, there are many intelligence applications for which they cannot be guaranteed. Often relevant information will be stated only once or may only be inferred and never stated explicitly. Furthermore, the results of the most recent TREC question-answer competition suggest that deep reasoning systems may now have reached a level of sophistication that allows them to surpass the performance possible using surface-based approaches. In the 2002 TREC competition, the POWER ANSWER system (4), which converts both questions and answers into propositional form and uses an inference engine, achieved a confidence weighted score of 0.856, a substantive improvement over the second placed exactanswer (5), which received a score of 0.691 in the main questionanswering task.A key component in the performance of the POWER ANSWER system is its use of the WORDNET lexical database (6). WORDNET provides a catalog of simple relationships among words, such as synonymy, hypernymy, and part-of relations that POWER ANSWER uses to supplement its inference system. Despite the relatively small number of relations considered and the difficulties in achi...