The use of memory in text processing

The performance of a natural language processing system should improve as it reads more and more texts. This is true both for systems intended as cognitive models and for practical text processing systems. Permanent long-term memory should be useful during all stages of text understanding. For example, if, while reading a patent abstract about a new disk drive, a system can retrieve information about similar objects from memory, processing should be simplified. However, most natural language programs do not exhibit such learning behavior. We describe in this article how RESEARCHER, a program that reads, remembers and generalizes from patent abstracts, makes use of its automatically generated memory to assist in low-level text processing, primarily involving disambiguation that could be accomplished no other way. We describe both RESEARCHER's basic understanding methods and the integration of memory access. Included is an extended example of RESEARCHER processing a patent abstract by using information about several other abstracts already in memory.