Using Memory in Text Understanding
It is clear that text processing proceeds at many levels simultaneously (Charniak, 1983; Marslen-Wilson, 1975; Schank, Lebowitz, & Birnbaum, 1980). Such processing presumably includes the access of detailed, long-term memory for the purpose of finding information relevant to a new text. It seems plausible that such high-level information should be useful in assisting low-level processing. We have suggested in our earlier work, as have others, that memory access may help determine resource allocation and in identifying the important parts of a text (Lebowitz, 1981; Schank et aI., 1980). However, these are imprecise ideas, and difficult to apply. For example, IPP (Lebowitz, 1980; Lebowitz, 1983a), a program that reads, remembers, and generalizes from news stories about international terrorism, might know from accessing memory that the destination of a
hijacking in a story that begins, "A United 727 en route to Miami was commandeered ... " is likely to be Cuba. However, it would not know how to use this information should the story continue "to Havana." Because of the inherent redundancy of language, it is easy to process this information bottom up. (Had no destination been mentioned, we would have used Cuba as a default.)
It is our feeling that the best way to use specific information in memory for text understanding in the context of current systems is to identify specific points during processing where memory can be applied. In this chapter, we present suggestions for using memory for low-level text processing in the context of RESEARCHER, a program that reads patent abstracts and builds up a generalized long-term memory (Lebowitz, 1983b). We illustrate this process with a version of RESEARCHER that simulates memory access by asking a user questions at key points during the processing of an abstract. (Answering these questions has been largely automated, but the memory search is not central to this paper.) Crucial to the development of RESEARCHER is that its text understanding process be robust enough to handle many patents without special preparation. We feel that memory application is necessary to achieve th~s ability.