Representing Complex Physical Objects in Memory

Researchers an artificial inteJ/igenc:e have proposed and implemented several representation systems for use in computer programs that "understand" natural language input. Notice4bly lacking from these 8y8tems is a rOOust and concise method of representing complu physical objects. This paper describes a langu.ageindependent scheme for encoding real-world objects in a manner that captures elements of how people describe them. Two m4jor groups of objects are distinguished: ''unitary'' objects that (Ire des crib" by a single "shape-ducriptor"; and "composite" objects th4t are represented by a frame-b4sed syatem that focusu on the physical rei4tions th4t mst among obiects. The heart of this scheme is a primitive-based framework that classifiu physical relations into three fundamental categories with five possibk properties. Our current work on RESEARCHER, a program that employs this scheme while reading patent abstracts, is also discussed.