Generalizations Based on Explanations

Abstract

phenomenon under study is a kind of " i n s i g h t l ea rn i ng " of procedural schemata. The system described here is designed to grasp some p r i n c i p l e under ly ing a na tu ra l language i n p u t. The under ly ing p r i n c i p l e r esu l t s in a new schema fo r the system. Once acqu i red, the schema serves the same purpose as the other schemata in the system: i t a ids in processing fu tu re na tu ra l language i npu ts. The neu t ra l term "schema " ra ther than " f rame" (Minsky (1975), Charniak (1976)) or " s c r i p t " (Schank A Abelson (1977)) is used to r e fe r to knowledge chunks because a frame (which is used to descr ibe s t a t i c ob jec ts as we l l as progressions of world s i t u a t i o n s) is too general a n o t i o n, and the no t ion behind a s c r i p t is- too s p e c i f i c. Sc r ip t s cannot be made to f i t a novel s i t u a t i o n nor are they designed to represent the more abs t rac t concepts tha t t h i s system w i l l need. The process that the system uses is ca l l ed explanatory schema a c q u i s i t i o n. The basic idea behind i t is that the causal connections in an understood representa t ion of a new input can be used to propose and propagate cons t ra in t s on s l o t f i l l e r s. That i s, from one p a r t i c u l a r instance of a s i t u a t i o n the system can "reason out " the a general s t r uc tu re under ly ing that ins tance. The system is there fore capable of learn ing from j u s t one example. There has been much work in the past on learn ing systems which s t a r t from a s ta te of very l i t t l e or no i n i t i a l knowledge ( f o r example, Minsky and Papert (1969), Meisel (1972) ). To a large ex ten t, these systems f a i l e d to l i v e up to the expectat ions of researchers. More recen t l y, l ea rn ing and knowledge a c q u i s i t i o n systems have been constructed on a f i r m foundat ion of r i c h