Adaptation-Based Explanation: Extending Script/Frame Theory to Handle Novel Input

1 Introduction

The ability to develop hypotheses to explain new, unex­ pected experiences is a hallmark of human intelligence. It is also a crucial concern within artificial intelligence, since intelligent computer systems need to construct ex­ planations in order to guide learning (as discussed in [De­ Jong, 1983] and [Mitchell et al., 1986]), to recover from planning failures (as discussed in [Hammond, 1986]), and to make sense of the stories that they read.

The hard· part of processing interesting stories is not understanding the language as such, but developing cre­ ative hypotheses about what might be going on in the story. For example, consider the following story: "Swale was a star three-year-old racehorse . Swale won the Bel­ mont Stakes. A few days later, he died." This story is an interesting one even though it is short and sim­ ple, because it gets a curious reader to start wonder­ ing what might have happened to Swale. The goal of the Adaptation-Based Explanation project (henceforth, ABE 1 ) is to build a system that can develop creative hypotheses about events, like the one described in the Swale story, by adapting the explanations that it already has in memory to new situations.

·This work was supported in part by the Defense Ad­ vanced Research Projects Agency, monitored by the Office of Naval Research under contract N0014-85-K-0108 and by the Air Force Office of Scientific Research, under contracts F49620-88-C-0058 and AFOSR-89-0100.

2 Inference is the issue

The principal issue in designing a computer program that builds explanations is how to bring the system's causal knowledge to bear on a problem so that it can efficiently infer an unseen cause of observed events.

Approaches to this inference problem have tradition­ ally fallen into two principal categories. The first ap­ proach was to build systems that had a large base of inference rules and were able to build explanations by chaining these inference rules together on the fly, in re­ sponse to each new problem. Examples of this approach include Rieger's [1975] conceptual inferencer, MYCIN [Shortliffe, 1976] and Wilensky's [1978] PAM program. Researchers who take this approach are led to concen­ trate on methods of controlling the combinatorial explo­ sion inherent in the inference-chaining process. Paral­ lel methods of rule activation, bi-directional search, and application of either domain-dependent or abstract, the­ matic search-control knowledge are among the important elements in this camp's bag of tricks.

On the other hand, another school of research seeks to view the entire explanation process as an index­ ing problem. These researchers postulate the existence of ready-made knowledge schemas, such as scripts or frames ([Schank and Abelson, 1977], [Cullingford, 1978], [Minsky, 1975], [Charniak, 1977]) that will provide all of the proper inferences if only the right structure can be found in memory. This is the "retrieve and apply" school of explanation construction. For script appliers, the inference problem is reduced to matching input from the story against a schema in memory This makes script appliers very efficient, provided that the stories they are faced with match their schemas. The concerns of the "retrieve and apply" school naturally revolve around how these knowledge structures are represented and es­ pecially around how they are organized in memory so that the right one can be found at the right time.

1 ABE is a descendant of SWALE, which was the joint work of the author, David Leake, and Chris Owens.

Both sides have strong arguments against the other, and both are right. The script-application approach originally arose out of the realization that the "chain­ ing + control" method was ignoring issues of memory. The systems it produced treated each experience as com­ pletely novel, and this limitation doomed those systems to perform expensive analyses every time they recognized a new instance of a plan. On the other hand, the script­ application approach can only be employed when mem­ ory contains a script that precisely matches the episode being processed. In order for the matching process it­ self to be relatively inference-free, it is important that the expectations represented in a script be at a very specific level. But this means that a "retrieve and ap­ ply" system is stymied when confronted with situations that stray even slightly from its specific expectations. In short, "chaining + control" is a weak approach be­ cause it knows nothing of stereotypes, and "retrieve and apply" is limited because it knows only its stereotypes. What's needed is a system that knows stereotypes, but also knows how to go beyond them when necessary.

3 Adding adaptation to script/frame theory

ABE is an extension of script/frame theory for handling input that is less stereotypical. The key to building that more flexible understander is to relax the assumption, which script theory relied on, that schemas will be re­ trieved and applied in situations that precisely match the situations that those schemas were built to describe. Instead one must assume that schemas will be retrieved in situations that are only sort of like those they were originally built to handle.

Broadening the range of situations in which a schema will be applied introduces the need to evaluate the newly­ produced explanations for problems, and to do some cre­ ative adaptation to fix any problems that are found . In an adaptation-based theory of explanation much of the burden of producing an appropriate explanation is moved out of the schema-retrieval module into the adap­ tation module (which we call the tweaker) . The emphasis is shifted from pulling a perfectly appropriate structure out of memory, to being able to work with whatever the best structure pulled out memory happens to be. Since this adaptation process is more time-consuming than simply applying structures straight out of the schema library, it makes sense to add a storage module as well, allowing the system to save and re-use the new variations of its structures that the adaptation module produces.

In the augmented theory, "retrieve and apply" evolves into "retrieve, apply, evaluate, adapt, and store". (Actu­ ally, the application, evaluation and adaptation steps are contained in a loop; when a new variation is produced, that new variation is instantiated and evaluated again. If there are still problems with the explanation it can be further adapted.) This is essentially an application of the case-based methodology to the problem of constructing explanations. (Other CBR-related adaptation work that has influenced my thinking includes [Hammond, 1986] and [Kolodner et al., 1985]. Some other work on adapt­ ing explanations includes [Koton, 1988] and [Simmons, 1988].)

Because ABE draws on schemas as a starting point, but can then make sensible changes to those schemas, it requires both a schema library and a knowledge base of facts and plausible inference rules to represent its do­ main knowledge. The schema library contains the vari­ ablized explanations (called Explanation Patterns, or XPs [Schank, 1986]) that the system will adapt. The XPs are inference networks (built out of the raw material from the knowledge base) that lead from the anomalies they explain back to possible causes. XPs can be thought of as augmented scripts which represent not just the surface expectations to be matched, but also the causal relation­ ships between the various expected events. This causal annotation aids in both the evaluation of instantiated explanations, and, as we shall see, in their adaptation.

The rest of this paper will concentrate on how the tweaker works. We will assume the existence of a re­ triever, evaluator, and storer, but won't describe how they operate. See [Owens, 1988] and [Leake, 1988] for some discussion of issues ·related to these tasks. We'll start by presenting some examples of the sorts of expla­ nation failures that the tweaker can handle, next move to a discussion of some of the tweaking strategies used to fix those failures, and then move to more general dis­ cussion of the mechanics of tweaking.

4 Explanation failure

The tweaker is initially fed a "failed" explanation (i .e. one that the evaluator rejected as inappropriate in the current situation) along with a problem description (ab­ breviated hereafter as XP-FD, for XP-Failure Descrip­ tion), indicating why the evaluator decided that the XP needed tweaking. There are many different ways that explanations can fail, and there are several ways to fix most failures. A tweaking strategy is an algorithm for searching the knowledge base to find substitutions, gen­ eralizations, or specifications that are needed to fix a particular kind of failure. For example, if the problem with an explanation is that it involves some actor per­ forming an action that he is incapable of performing, then the failure may be fixed by substituting in either a new action that the old actor could perform or a new actor who could perform the old action. Furthermore, there are several ways to search for prospective substi­ tutions. What follows are a few examples of the main categories of tweaks that system handles.

4.1 Plausibility failures

Plausibility failures correspond to explanations that do not make sense because they contradict some aspects of the explainer's world-model, as represented in the knowl­ edge base. Consider the following examples of plausibil­ ity failures :

• The Bias Story: College basketball star, Len Bias, died one day after being drafted by the Boston Celtics. He was the first pick in the NBA draft .

Jim Fixx XP: Someone who regularly engages in recreational jogging also has a hereditary heart de­ fect. The stress on the heart from exertion caused by jogging combines with the defect to cause that person to have a heart attack and die.

Failure: Len Bias wasn't known as a recreational jogger. (An actor of some action in the explanation is not known to perform that action.)

Fixing the failure: By substituting playing bas­ ketball for recreational jogging, the adapter can build an excellent explanation based on this XP.

• The Swale story: Swale was a star three-year­ old racehorse. The day after winning The Belmont Stakes, Swale died.

Spouse insurance XP: Someone who is greedy and doesn't really love his/her spouse kills the spouse in order to collect the life insurance money.

Failure: Swale has no spouse and no life insurance. (A slot referenced in the XP does not actually exist).

Fixing the failure: By searching for other ac­ tors and similar motivations, the adapter can build from this explanation the reasonable hypothesis that Swale's owner got greedy and killed him in or­ der to collect the property insurance.

Plausibility failures can arise from explanation that involve mismatched actor-action pairs, mismatched implement-action pairs, mismatched object-action pairs, reference by the explanation to missing slots, tempo­ ral sequence problems, and specific contradictions in be­ tween beliefs in XP and beliefs in the system's knowledge base.

4.2 Vagueness failures

Vagueness failures correspond to explanations that are not detailed enough, not sufficiently convincing, or do not contain the kind of information that suits the ex­ plainer's needs. What counts· as sufficiently convincing or sufficiently detailed is a function of what the expla­ nation will be used for.

For example:

• The Pan Am Story: Pan Am flight 103, en route from London to NYC, exploded in mid air, killing all aboard.

Terrorist bombing XP: Someone who is engaged in intense political conflict with the people of a par­ ticular nation may kill citizens of that nation by planting bombs in crowded areas in that nation.

Failure: Does not specify enough about who did the bombing. Note that this might not be a prob­ lem if the understander were a Pan Am engineer who just wants to know whether it was a design flaw that caused the crash, but it would be a prob­ lem if the understander were the government agency responsible for retaliating against the perpetrators. (A slot-fillet is insufficiently specified.)

Fixing the failure: By searching for a more spe­ cific description of someone who might have had appropriate motivation to perform the action, the adapter can conjecture that perhaps an Iranian ter­ rorist planted the bomb on the American Pan Am jet in order to retaliate for America's destroying an Iranian airliner.

• A Suicide Bomber Story: A teenage girl ex­ ploded a car bomb at a joint post of Israeli troops and pro-Israeli militiamen in southern Lebanon, killing herself and a number of Israeli soldiers.

Terrorist bombing XP: Someone who is engaged in intense political conflict with the people of a par­ ticular nation may kill citizens of that nation by planting bombs in crowded areas in that nation.

Failure: This doesn't explain an important part of the anomaly -why someone would do something that resulted in her own death. (A decision that the explanation claims occurred is not sufficiently motivated by the XP.)

Fixing the failure: One way to make an action in an XP seem more motivated is to add an expla­ nation of why the negative side effects of the action would be less important than expected to the actor involved. By employing this strategy the adapter can hypothesize that in addition to having the above political conflict motivation, the bomber was termi­ nally ill, and therefore did not value her own life as highly as most people would.

Vagueness failures mainly arise when the XP contains insufficiently specified slot-fillers, or beliefs that, while not contradicted by anything in the knowledge base, re­ quire more causal support.

5 Adaptation Strategies

In the previous section the idea was to describe the kinds of problems that the explanation tweaker is faced with, and to give a feel for the kinds of output that it pro­ duces. The heart of the adaptation process is a library of tweaking strategies, each of which is a program that is capable of fixing a class of explanation failures. For any given explanation failure that the tweaker can han­ dle there will be a set of tweaking strategies that apply. The tweaker ranks this set according to how efficient the strategy has been in the past, how often it has worked in the past, and how closely tailored it is to the failure at hand. It then runs the strategies in best-first order until one of them produces an acceptable explanation. Now let's examine some of the strategies.

5.1 Substituters

Substituters attempt to fix plausibility problems by re­ placing a slot filler in one of the beliefs of a failed XP with a component that could have the same causal con­ sequents. For example, one of the strategies available for fixing actor-action mismatches like the one that occurred with the Bias story above is:

Tl: Replace old action with an action stereotypically as­ sociated with the actor.

This is the rule invoked to change from recreational jogging to playing basketball in the Bias example above. The idea is to climb the actor generalization hierarchy to find the categories that the actor is in (such as bas­ ketball player and college student for Bias), and then to follow actor-theme links to find the actions associated with those categories. (for example, college students at­ tend classes; basketball players play basketball). These are the candidates for substitution.

Next, the actions that are candidates must be exam­ ined to see if they link up with the original XP; are there inference rules that indicate that the new action could have caused the same things that the old action caused? Playing basketball involves running, just as recreational jogging does, so it is a good substitution to actually try. Attending classes doesn't involve running, or physical exertion, or anything in the original chain leading to death, and therefore is not a good substitution to try. The causal links within the original XP are crucial to making this distinction.

Briefer descriptions of some other substitution tweaks follow:

• T2: Replace old action with an action closely related to it in the action hierarchy.

For example, recreational drugs is a kind of drug taking, and a kind of recreational activity. So, if recreational drugs shows up somewhere that it isn't appropriate, consider other kinds of drug taking (such as medicinal, or performance-enhancing) and other kinds of recreational activities.

• T3: Replace old action with an action indexed as causing one of the events in the XP.

For example, if recreational jogging couldn't have been involved in the heart attack, consider other stereotypical causes of heart attacks, such as a sud­ den scare, or a bad diet.

• T4: Replace old actor with an actor known to have mo­ tivation mentioned in XP.

For example, Swale didn't have a spouse, but his owner might have stood to collect money upon his death.

• T5: Replace old actor with an actor who old actor could have made perform old action.

For example, if the explanation says that Swale's owner killed him, but the owner was known to be somewhere else at the time, maybe it was one of the owner's employees.

5.2 Generalizers

Generalizers fix plausibility failures and XP application failures by producing a version of an explanation that applies to a broader class of situations, at the cost of removing some of the detail from the hypothesis. There are two ways that this can be done:

5.2.1 Component generalizers

Component generalizers work by altering a particular slot filler in one of the beliefs of the XP. In this sense they are similar to substituters, except they try to gen­ eralize the invalid component rather than search for an alternative. Examples:

• T6: Generalize old action to make it compatible with new actor.

For example, go from recreational jogging to physi­ cal exercise.

• T7: Generalize constraint on actor to make it compatible with current actor.

For example, change an XP that calls for a Moslem religious fanatic to simply call for any religious fa­ natic to handle an example where the actor was known to belong to another religion.

5.2.2 XP simplifiers

XP Simplifiers generalize the structure of the XP by make a more global change than those which simply alter one of the components.

T8: Delete problematic belief

Removes conflicting beliefs from the explanation. The resulting explanation thus contains less infor­ mation, but may still be useful in many contexts.

5.3 Specifiers

Specifiers fix vagueness problems by finding detail to add to an explanation, making it less general, but richer in in­ formation content. There ·are two sub-categories of spec­ ifiers, roughly analogous to the sub-categories of gener­ alizers:

5.3.1 Component specifiers

Component specifiers add detail to the explanation by making the description of some slot-filler more specific. Examples:

• T9: Specify actor description by finding an actor which matches that description, and that is also mentioned elsewhere in the XP or elsewhere in the story being processed.

For example, in applying the Spouse Insurance XP to Len Bias, the idea that the Celtics killed Bias for the insurance money represents an off-the-wall, but rather creative hypothesis that can be generated by this tweak.

• TlO: Specify actor description by finding an actor match­ ing description that has motivation mentioned in the XP.

For example, Iran declared an intention to retaliate for the downing of one of its airliners, so it is known to be motivated to destroy American airliners, such as Pan Am 103.

5.3.2 XP elaborators

XP elaborators add detail to an explanation by build­ ing more structure into it. These strategies do things like add causal links between beliefs in the explanation, or add additional explanation to support one of the be­ liefs, or splice two XPs together to form a more complete explanation. Elaborators can involve recursive calls to the explainer. Examples:

• Tll: Add to the causal connections between two beliefs by splicing in another XP that explains how one would cause the other.

• T12: Explain a decision made by one of the actors by splicing in an explanation of why the good effects of the decision would be more important to the actor than to most people. For example, killing the Israeli soldiers might have become extremely important to the suicide bomber if they had killed someone in her family.
• T13: Explain a decision made by one of the actors by splicing in an explanation of why the bad effects of the decision would be less important to the actor than to most people. For example, dying might be less important to the suicide bomber if she was convinced that she would be rewarded after death, or if she were terminally ill and knew she would die soon anyway.
• T14: Explain a decision made by one of the actors by splicing in an explanation of why the actor might not have known about a bad effect of the decision. For example, perhaps the people who convinced her to do this had not told her that the car would ex­ plode while she was inside.

6 Overview of the tweaking process

Tweaking an explanation involves two steps, selecting a tweaking strategy to try, and then executing that strat­ egy. Both of these tasks, tweak selection and tweak execution have a number of sub-steps. The overall idea is to retrieve a set of tweaking strategies suggested by the XP-Failure Description that was generated by the explanation evaluator, to make an estimate of which of the retrieved tweaking strategies is the most promising in the current situation, and then to figure out how to apply that most promising tweaking strategy to the XP in question, using the knowledge in memory to make the kind of changes to the XP that the strategy calls for .

6.1 Tweak selection

The first job a tweaker has when it is called upon to fix a failed explanation is to decide which XP Tweak­ ing Strategy to apply. There are three sub-stages to the tweak selection problem: tweak retrieval, tweak filtering, and tweak ranking.

Tweak retrieval: Tweak retrieval involves pulling a first pass set of tweaking strategies out of the system's library of tweaking strategies. The tweak library of is organized in a straightforward fashion -each strategy is indexed under a variablized XP-FD pattern (abbreviated TIP, for Tweak Index Pattern) that corresponds to the type of failures that the strategy addresses. A TIP is a variablized structures that indexes a tweak in the tweak library. TIPs are constructed by the programmer and are static. An XP-FD is a structure that describes a specific explanation failure. XP-FDs are created dynamically by the evaluator. Each TIP matches a class of XP-FDs. The tweak retrieval step simply involves collecting all the strategies in the library whose TIP matches the XP­ FD being addressed. This search process is currently implemented using a variablized discrimination net with the patterns serving as indices and the current failure as the key.

Tweak filtering: Not all the strategies that match a given XP-FD will be applicable to the XP at hand. For example, T4 (described above), which involves substi­ tuting another actor mentioned in the XP could not be used to make the FIXX XP applicable to Bias, since the XP mentions no other actors. Each tweak has a set of constraints associated with which get run on the current XP to determine if it makes any sense to try that tweak on that XP. If and of the checks fail, the tweak is filtered out from the list of candidates.

Tweak ranking: Not all the strategies which don't get filtered out are equally likely to praduce desirable results. Therefore, tweaks are tried in best-first order, with ranking based on the following three criteria:

• Match Specificity is a measure of how closely tai­ lored the tweak is to the XP failure at hand. For example, T8 can apply to just about any plausi­ bility problem by removing the implausible belief altogether. This very general strategy often leaves the explanation much weaker, however. since it reduces the support for any conclusions that pre­ viously depended on the deleted belief. Closely­ tailored strategies are preferred over more general ones. In the current implementation this calculation is done by a simple method - counting the number of constants that match between the XP-FD and the tweak's TIP. A strategy with a more detailed TIP will thus match in fewer situations but will be more highly ranked when it does match.
• Cost Estimate is a measure of the amount of time that a tweak is likely to consume. For example, a tweak that simply deletes an offending belief is very fast, while a belief that searches memory for an ap­ propriate generalization is medium expensive, and a tweak that builds an entire sub-explanation to sup­ port a premise in an XP can be very expensive. In the current implementation, the "cost" of a tweak is estimated based on a very simple-to-calculate num­ ber - the average number of inferences that this tweak has performed when run in the past.
• Hit Ratio is a measure of how often a given tweak has produced explanations that the evaluator has found usable in the past. When the tweak produces an explanation that the evaluator rejects its hit ra­ tio goes down, when it produces an explanation the evaluator decides to use, the hit ratio goes up.

Explanations are needed for different reasons at dif­ ferent times, and this affects which tweaking strategy is most appropriate in a particular situation. A system in which ABE is embedded can control the tweaker's behav­ ior by determining the weight that each of the above fac­ tors should be given relative to one another. For exam­ ple, in situations where getting some hypothesis quickly is more important than necessarily getting the most de­ tailed or most plausible explanation possible, the cost estimate could be weighted more highly than the other two factors.

6.2 Tweak execution

A crucial issue that a theory of adaptation must an­ swer is what level of knowledge belongs in the adaptation strategies. One school of thought is to make the level of knowledge very general, sticking to rules like: When an explanation provides an inappropriate slot-filler, replace it with another slot-filler. This approach is elegant in that it produces a small number of domain-independent strategies, but it doesn't promise much in the way of effi­ ciency because the strategies don't specify how to search for the appropriate replacement. On the other hand, one could adopt very domain-specific adaptation rules that say things like: When looking for someone who killed a racehorse, consider the horse's owner. Such rules pro­ vide excellent guidance to the search process once se­ lected, but there will be very many of them, and one is left without much of a general theory of tweaking. ABE's adaptation rules fall somewhere in between. They are more problem-specific than the general rules described above, but they are still quite domain-independent. The knowledge that is coded into them is not domain knowl­ edge - that is coded into the knowledge base. Instead, what tweaking strategies know is how to search that knowledge base in order to make the necessary modi­ fications.

A tweak is defined by three things: the part of the XP that it modifies, the method of searching for the do­ main knowledge necessary to do the modification, and a set of checks that have to be run on the modifica­ tions to see if they make any sense. For example, Tl and T2 are candidates for fixing the same sets of ex­ planation failures, and they both work by replacing the actor in one of the XP 's beliefs. However, they differ in the way that they search for the replacement (Tl uses the actor-generalization hierarchy and actor-theme links, while T2 uses the action-generalization hierarchy). Since they use different algorithms to search for substi­ tution candidates, the set of checks they must perform is also different. For example, T2 must check to see that any actions it considers are compatible with the original actor, while Tl does not since that is guaranteed by its search algorithm.

7 A tweak in action

Several, though by no means all, of the tweaking strate­ gies discussed above are implemented in the ABE com­ puter program. The program output that follows is in­ tended to demonstrate how one of the system's tweaks actually operates. The transcript has been heavily edited to meet space requirements, and concentrates on show­ ing the tweak execution phase. The system is attempting to fix a problem with the application of the FIXX-XP to Swale's death (Swale was not a recreational jogger). It searches for actions stereotypically associated with Swale. It finds three. One (competing in horse races) is the antecedent of an inference rule involving the run­ ning action mentioned in the original explanation. This rule forms the basis of a new explanation in which com­ peting in races substitutes for recreational jogging.

Running a tweak:

Substitute an action that is stereotypically associated with the actor.

Anoaaly to be explained: SWALE died. XP being tweaked: FIXX-XP

Probleas is with SWALE as the actor of DID RECREATIONAL JOGGING Problea type - Actor-action-aisaatch : physical-disability

3 actions associated with SWALE being considered:

SWALE COMPETED IH HORSE RACES BECAUSE SWALE is in category RACEHORSES AND RACEHORSES had theae COMPETED IH HORSE RACES.

Trying to link up to: ("SWALE ran") This chain DOES link up with the rest of the xp: Inference: ?X? COMPETED IH HORSE RACES -> ?X? ran

SWALE TOOK PERFORMANCE DRUGS BECAUSE SWALE is in category RACEHORSES AND RACEHORSES had theae TOOK PERFORMANCE DRUGS.

RACEHORSES had theme TOOK PERFORMANCE DRUGS

BECAUSE RACEHORSES is in category ATHLETES AND ATHLETES had theae TOOK PERFORMANCE DRUGS.

Trying to link up to: ("SWALE ran") Does HOT link up with the rest of the XP. Cannot create a tweaked explanation based on SWALE TOOK PERFORMANCE DRUGS

SWALE ATE OATS BECAUSE SWALE is in category RACEHORSES AND RACEHORSES had theme ATE OATS

RACEHORSES had theae ATE OATS BECAUSE RACEHORSES is in category HORSE AND HORSES has theme ATE OATS

Trying to link up to: ("SWALE ran") Does HOT link up with the rest of the XP. Cannot create a tweaked explanation based on SWALE ATE OATS

Creating a new tweaked explanation base on: SWALE COMPETED IH HORSE RACES

Adding an index to FIXX-XP-1: ?X? COMPETED IH HORSE RACES

Adding: ?X? ran BECAUSE ?X? COMPETED IH HORSE RACES. Adding: ?X? COMPETED IH HORSE RACES BECAUSE ?X? had theae COMPETED IH HORSE RACES. Adding: ?X? had theae COMPETED IH HORSE RACES BECAUSE ?X? is in category RACEHORSES AND RACEHORSES had theae COMPETED IH HORSE RACES. Deleting a belief: ?X? DID RECREATIONAL JOGGING

Deleting supporter inference:

Preaise: ?X? DID RECREATIONAL JOGGING Deleting supported inference: ?X? ran BECAUSE ?X? DID RECREATIONAL JOGGING. Deleting orphaned beliefs.

FIXX-XP-1.EXPL-1 Preaise: THE HEART OF SWALE is in category HEREDITARY DEFECTIVE HEARTS Premise: SWALE is in category RACEHORSES SWALE had theae COMPETED IN HORSE RACES BECAUSE SWALE is in category RACEHORSES AND RACEHORSES had theae COMPETED IN HORSE RACES. THE HEART OF SWALE vas veak BECAUSE THE HEART OF SWALE is in category HEREDITARY DEFECTIVE HEARTS. SWALE COMPETED IN HORSE RACES BECAUSE SWALE had theme COMPETED IN HORSE RACES. SWALE ran BECAUSE SWALE COMPETED IN HORSE RACES. SWALE had a VERY HIGH degree of PHYSICAL CONDITIONING BECAUSE SWALE had a VERY HIGH degree of EXERTION. SWALE had a VERY HIGH degree of EXERTION BECAUSE SWALE ran. SWALE had a heart-attack BECAUSE THE HEART OF SWALE vas veak AND SWALE had a VERY HIGH degree of EXERTION. SWALE died BECAUSE

SWALE had a heart-attack.

8 Conclusion

The goal of this paper has been to describe and to advo­ cate an adaptation-based theory of explanation. I am at­ tempting to extend script/frame theory by using adapt­ able explanation patterns as my knowledge schemas, and by providing my system with the means of adapt­ ing its stored schemas to new situations. I have devel­ oped a taxonomy of explanation failures and an arsenal of failure-specific, but domain-independent adaptation strategies.

ABE learns new variations of its schemas as it pro­ cesses new stories. It attains the efficiency advantages of a script-applier and the flexibility of a system that constructs explanations from scratch.

Acknowledgments

I would like to thank Chris Owens, Louise Pryor, and the IJCAI referees for their very helpful comments on drafts of this paper.