Generalizations Based on Explanations

GENERALIZATIONS BASED ON EXPLANATIONS

Geralrl DeJon~

Coordinaterl Science Laboratory University of Illinois 1101 West Sprin~field Avenue Urban~, IL 61801

This paper describes a new pro.iect at the University of Illinois in computer learnin~. The phenomenon under sturly is a kind of "insi~ht learning" of procerlural schemata. The system described here is designed to ~rasp some principle un<ier lying a natural l·mgu.'lge input. The underlying principle results in a new schema for the system. Once acquired, the schema serves the same purpose as the other schemata in the system: it aids in processing future natural language inpute.

The neutral term "schema" rather than "frame" (Minsky (1975), Charniak (1976)) or "script" (Schank & Abelson (1977)) is used to refer to knowled~e chunks because a frame (which is used to describe static objects as well as progressions of world situations) is too general a notion, and the not ion behind a script i~ too specific. Scripts cannot be made to fit a novel situation nor are they designed to represent the more abstract concepts that this system will neer!.

The process that the system uses is called explanatory schema acquisition The basic irlea behinrt it is that the causal connec tions in an unde rstoort representation of a new input can be user! to propose and prop~gate constraints on slot fille rs. That is, from one particular instance of a situation the system can r eason out" the a ~eneral struc ture underlyin~ that instance. The system is therefore cspable of l ea rning from just one exsmple.

There has been much work in the past on lea rning systems whi ch start from a state of very little or no initial knowledge (for example, Minsky and Papert (1969), Meisel (1972)). To .a large extent, these systems fa i led to live up to the expectations of researchers. More recently, learning and knowledge acquisition systems have been constructed on a firm foundation of rich initial knowledge (Lenat (1976), Soloway (1977), Buchanan and Mitchell (1978), McDermott (1979), Se lfridge (1980)).

'l'hte wo r·k reported in this pap"'r wqs O'ltlpport"!l in pa rt by thte Offi ce of Naval R~s e ·rch unjer conrqct N000 14-7~-C -06 1 2 .

The system described here is a knowlerl~e baser! one. It must cont~in much background inform'ltion on how the world behaves and on the goals and motivations of people. From this, it is able to

learn more about the behavior of the world.

In the example below assume the system ioes not yet have schemata for kidnap or extortion. It does, however, possess a considerable quanti t y of background information about stealing, bargainin~, the use of normal physical objects, anrl goal s of people, companies and institutions. Some of this knowledge is in the form of other schemata already built into the system.

Paris police rlisclosed Tuesrlay that a m'ln who i<ientified himsel.f Jean Maraneux abducterl the 12 year old daughter of wealthy Parisian businessman Michel Boullard late last week. Boullarrl received a telegram rlemanrling that 1 million francs be left in a lobby was te basket of the crowded Pompirlou Center in exchange for the girl. Asking that the police not intervene, Boullard arranged for the delivery of the money. His dau~hter was found wandering blindfoliei near his downtown office on Monday.

In processin~ the story the system ·.;ill be able to learn a general schema for processing kirlnap stories from this one instance. Using explanatory schemata acquisition, understanding is viewed as the process of explaining input events. The explanations can then be used to generalize a single event into a new schema.

In processing this example without a KIDNAP or EXTORTION schema the system cannot explain all of the events via existin~ schemata. In particular, existing schemat'l cannot explain why Maraneux might steal Boullard's daughter. While this is quite clearly an instance of takin~ something that belongs to someone else, there is no motivation for it. The daughter has no apparent value to Bou ll.'l rd, and a person, un 1 ike money, canna t be usei to acquire other V'llued goods. The system r~quires motivations for major volitional 'lctions ( su~h SO'! a character invoking the STEAL schema). TIJerefore it is confusei 'lt this point.

The confusion is resolved by the next sentence. This input invokes the BARGAIN schem~. The system understan~s the motivation for Maraneaux to initiate the bargaining event: he is trying to acquire money which it knows to be 3 possible ~oal of any human. Furthermore, this provides the motivation for the STEAL event. ~araneaux use~ the STEAL schema to satisfy the precondition of the BARGAIN schema of possessing the item to be trade~ .

Resolving the confusion causes the system to invoke its explanatory schemata acquisition procedure. This procedure does two thin~s. First, it constructs a new schema composed of a STEAL event and a BARGAIN event where the STEAL is used to satisfy a precondition of the BARGAIN. Second, constraints on the slots for the new schema are derived from the knowledge in the systems STEAL and BARGAIN schemata and the story as follows:

1. the slot filled by 1 million francs is generalized to be any amount of money.
2. the slot fille~ by Maraneaux (the kidnapper) is generalized to be any adult human.
3. ') the slot filled by the daughter is generalized to be anyone with close personal ties with (4) .
4. the slot filled by Boullard is generalized to be any human who both has the amount of money to fill (1) and a person with close personal ties to fill ( 3) .

Thus the system now has a schema that can be used to process a new story about a person stealing another person in order to trade him back for money. This is, of course, a frrst approximation to a KIDNAP schema.

One might make the argument that learning the KIDNAP schema is unnecessary since the system could, after all, correctly process the example input without it. Why not simply rely on the more general planning/goal knowledge and forget about the schemata? There are three re3sons.

efficiency can rely on goal following to do. First, there is an The more a system schemata, the less planning it will have considers tion. pre-package.i and complex

Second, planning systems have proved to be more fragile and dis tine tly less successful th::m script/frame/schema systems. It might well not be possible to build a successful planning type system to deal with .an interesting domain. The problem is th~t the fr3gility of ~ plqnner is multipliod by Lon~ reasoning chains, a domain rich in alternatives (as the real world tends to be) ~nd the need to invoke the planner many timas in the course of processing an input. In the proposed system, most of the processing is handled by the schem~ knowledge. Only when a new schema is being constructed or refined must the system resort to less reliable knowl edge sources.

Finally , script/frame/schema systems have been much criticized as 'ad hoc' . This is due to the fact that programmers invariably buili in all of the necessary (and no unnecessary ) knowledge structures . A system that could acquire these knowledge structures for itself would help dispel that criticism .

This is not the end of the learning process for the KIDNAP schema . The system is designed to refine and exp~nd initial schemata in the co~rse of processing more stories. We hope th~t the syst~m will bo> .~bl·~ to exp~n < l th·· initi·!l KIDNIIP >H:hcm·, into ~ general ~XTORT schema by recognizidg th~t other things besides people can be stolen ~nd tnl e l b<~ c k.

The system is only in the beginning st3ges. As yet it is not implemented.

In conclusion, it should be pointed out that this kind of learning does not depend on correl~tional evidence. That is, the new schema is not constructed by noticing the similarities and differences among a large number of inputs (like Winston (1970) or Fox & Reddy (1977)). Nor does it construct special purpose knowledge structures from mor~ general ones already existing within the >ystem (e .g., Lebowitz (1981)). R~ther, the system is capable of one tria l " learning not unl ike So lowqy's (1977). From just one kidnap story, th <~ system constructs a plausible KIDNAP schema. This schema can later be refirted and expAnded, but fro~ the beginning it is ~ viable new schema capAbl e of ai~ing future processing.