An Overview of the FRUMP System
Gerald DeJong Yale University
FRUMP (Fast Reading Understanding and Memory Program) is a newspaperskimming program developed at the Yale Artificial Intelligence Project to skim and summarize news articles. The program's input comes directly from the United Press International news wire. Thus FRUMP is one of the very few AI programs that attempts to deal with unrestricted real-world input. The program routinely works on inputs never before seen by either FRUMP or its programmers. FRUMP processes input stories on such diverse topics as reports of plane crashes, countries establishing diplomatic ties, forest fires, and wars. Past natural-language processing programs have restricted input text in one of two ways. Some ( Brown & Burton, 1975, Epstein & Walker, 1978; Winograd, 1972; Woods & Kaplan, 1971) severely constrain the input domain to a small subset of English. Others, for example Cullingford ( 1978), only concern themselves with a small number of "target inputs" to demonstrate the feasibility of their approach. FRUMP processes stories on a wide range of topics, and routinely produces correct summaries for stories for which it has not been primed.
The main thrust of FRUMP is that text analysis, even at the lowest levels, can benefit from the pragmatic knowledge of the system. It is not new to endow a natural-language processing system with large amounts of world knowledge. Indeed, the prevailing view is that the system must be able to reason and be capable of problem solving in order to understand natural language about a domain. For example, Wilensky's work on planning ( 1978) was motivated by the needs of a natural-language system; Charniak's frames ( 1977), developed to help process natrual language, work equally well for problem-solving applications.