From: AAAI Technical Report WS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Retrieval strategies for tutorial stories* Robin Burke Institute for the LearningSciences 1890 Maple Ave., Evanston IL 60626, USA burke@ils.nwu.edu suggestions, to namejust a few possibilities. This paper describes in detail three of the retrieval strategies that are used to produce SPIEL’s storytelling behavior. Each involves a comparisonbetweena story and the situation in which it might be told, a comparisonthat differs according to the needs of the educational context. SPIEL’sretrieval strategies select different subsets of features froma story’s index and use a variety of measurements of fitness including similarity, dissimilarity and other kinds of relations. Abstract Retrieving stories to present to students is a challenging application of case retrieval. This paper describes SPIEL, a system that retrieves tutorial stories, stored on video, for educational purposes. Although CBRmethods are employed in SPIEL, its task requires a different emphasisthan typically found in problem-solving CBRsystems. One of the most significant of these is the centrality of multi-purpose retrieval in educational storytelling. SPIELhas a set of storytelling strategies, corresponding to different educational roles that stories can play, such as providing counter-examples or projecting possible results. To find stories that can fill these roles, the system uses a variety of comparisons including similarity, dissimilarity, and other relations. This paper describes three of these strategies in detail, showinghowthe strategies function in retrieval, what kinds of knowledgethey use, and how they make use of SPIEL’sindices. What SPIEL does Introduction Is case retrieval a quick, associative process or is it a strategic, inferential one? Case retrieval can be viewedas a simple process that gathers raw material for the rest of the case-based reasoning process to use (Waltz 1989). However,there are manyuses for cases that require highly specific comparisonsdriven by strategic considerations. I have been investigating strategic case retrieval for educational purposes by building a program, SPIEL(Story Producer for Interactive Learning), to approximatewhat an instructor does in recalling a pedagogically-appropriate story to tell to a student. Tutorial storytelling involves using stories in a variety of educational roles. They can project possible results of students’ actions, provide counter-examples, and give * This workis supported in part by the DefenseAdvanced ResearchProjects Agency,monitoredby the Air ForceOffice of Scientific Research under contract F49620-88-C-0058 and the Office of NavalResearchunder contract N00014-90-J-4117, by ONRunder contract N00014-J-1987,and by the AFOSR under contract AFOSR-89-0493. The Institute for the LearningSciences wasestablished in 1989with the supportof AndersenConsulting, part of The Arthur Andersen WorldwideOrganization. The Institute receives additional support from Ameritech and NorthwestWater,Institute partners, and fromIBM. 118 SPIEL is designed to assist students who are learning social skills. It is embedded in an intelligent learning-bydoing architecture called Guided Social Simulation or GuSS. GuSS provides a social simulation in which students can safely practice social skills, such as those required by diplomacy or business. Currently, we are using this architecture to develop an application, YELLO, for teaching employees of Ameritech Publishing the fine points of selling Yellow Pages advertising. The goal in YELLO, and other GuSSapplications, is to accomplish for the social environment what the flight simulator accomplishes for the physical environment of the cockpit, letting the student learn by doing. For more about GuSS and its precursor ESS, see (Kass et al. 1993), (Kass Blevis 1991) and (Blevis &Kass 1991). Within GuSS, SPIEL is like an experienced instructor watching over the student’s shoulder. It monitors the simulation and presents stories from its library whenthey are relevant to the student’s situation. Telling stories in the context of simulation is a particularly useful way to connect the student with an expert’s experience. Stories help bring the simulation to life, and the student’s activity in the simulation helps makethe stories comprehensible. SPIEL’sstories are video clips of practitioners telling anecdotes about their ownon-the-job experiences. SPIEL’s knowledgeof its stories comes from manually-constructed indices entered into the systemto describe each story. The system currently has about 180 such stories gathered from interviews with Yellow Pages account executives. An earlier version of the program, described in (Burke &Kass 1992), contained 170 stories about selling consulting services. Standard problem-solvinl~ Story retrieval in SPIEL Incremental Beforeretrieval Cue composition Relrieval criteria Solves a similar problem Makesan educational point Case structure Represents problem solution Video clip No Yes Mandator~retrieval Yes No Case evaluation Between-case competition Yes No Table 1. Differences betweencase retrieval in _oroblem-solvingCBRand story retrieval in SPIEL Story retrieval vs. case retrieval Educational tasks emphasize different requirements for case retrieval than most problem-solving tasks. Table 1 summarizes some of the most important differences in emphasis. Cue composition: One of the major differences between SPIEL and the standard problem-solving CBRmodel (Kolodner & Jona 1992) is the way retrieval cues are put together. CBRsystems create a retrieval cue by analyzing a statement of the problem to be solved. SPIELretrieves its stories based on a continuous stream of actions by the student and the GuSSsimulation. Anyevent in the history of the interaction could be relevant to the retrieval of a tutorial story. So, SPIELhas to build its retrieval cues incrementally throughout this on-going process. Mandatoryretrieval: A case-based problem solver must retrieve something, otherwise there will be no basis for building a solution. If what is retrieved is not a directly-applicable solution, it can be adapted. However, in teaching, there are manystudent states for which there will be no appropriate story to tell. Becauseit cannot adapt its stories, SPIELhas to retrieve only closely-relevant ones. A story that is far afield from the student’s immediate concerns will be confusing. If SPIEL cannot find a very good story, it is better off waiting for the student to do something else. Usually, there is an appropriate story about once every 10-20 student actions. Case evaluation: In the standard CBRmodel, the retrieval step uses indices to retrieve a set of candidate cases. The candidate cases themselves are then examined and evaluated, and the best case is chosenas the basis for a problem solution. The evaluation step enables the system to reject inappropriately retrieved cases. SPIELhas no language capacity with which to understand or evaluate the video clips it retrieves. It must retrieve conservatively, because whatever is retrieved will be presented to the studenL Between-casecompetition: Educational stories are not in competition to be the single right answer, as in many retrieval models,such as (Thagardet al. 1990). If there are stories from experts with opposing viewpoints on the student’s situation, SPIELneeds to showthe student the whole range of opinion by making all retrieved stories available. These differences have three notable consequences for the designof a tutorial case retriever. 119 Because SPIEL does not have access to the content of what it is retrieving, the index has to contain everything that the system will need to decide to tell the story. It needs indices that are more detailed and complex than those typically used in case retrieval. There is an implicit theory of problem solving in the similarity metrics found in problem-solving CBR:the more similar the input problem is to the problemsolved by the case, the morelikely it is that the solutions will also be similar. SPIEL’sretrieval strategies also embodya theory of good cases for the educational context. They involve different kinds of judgmentsfrom those found in similarity metrics. In SPIEL,a tutorial opportunity is defined as a situation for which there is a relevant story to tell. It does a storyteller no goodto recognize a goodtime to tell a story it doesn’t have. A similar insight was behind the design of ANON(Owens 1990), which used its case base determine what features to search for in the input problem. SPIEL uses its story base in a similar way to guide the search of the events in the simulation. The rest of this paper focuses mainly on retrieval strategies, but I touch briefly on the indexing and implementationissues to put the strategies in context. Implementingstorytelling strategies One way to think about the problem an educational storyteller faces is to think of each story as a possible lesson and each strategy as a way to teach it. Storytelling strategies indicate the conditions under which the goal of telling a story can be achieved. Since a storyteller has manystories, any one of which could be relevant at any time, it is useful to think of it as an opportunisticsystem. However,educational storytelling is not as open-ended as the general case of opportunism whose complexities were laid out in (Birnbaum 1986). SPIEL’s storytelling strategies comparestories about an activity to specific contexts of student action. They seek coherence and relevance, not distant analogies. It is possible to describe concretely what an opportunity to tell a story using a particular strategy wouldlook like. SPIELalso has a strong advantage over the problem of opportunism in that it is embeddedin a simulation. No truly novel actions can occur in GuSSsince the student is constrained by the program’s interface and the simulation operates in known ways. SPIEL knows with certainty what actions can and cannot happenin its world. Storage generation ~ IMic~~ St°rytelling I strategies Opportunity-recognition rules Retrieval C Opportunity-recognition rules Retrieved Eventsfrom---{~1 Marcher ~- story/strategy simulation I I Language generat~ Story with bridge andcoda Figure 1. SPIEL’sprocessing SPIEL’s problem of opportunism is therefore much simpler than the general one. The simulated world provides a limited space in which events can occur; storytellingstrategies single out preciseareas of the space that constitute opportunities. Theseproperties enable SPIELto use what Birnbaumcalls the "elaborate and index" modelof opportunism: ...spend someeffort, whena goal is formed, to determinea numberof situations in whichit mightbe easily satisfied...and then indexthe goal in termsof all the features that mightarise in suchsituations. (pg. 146) SPIELworksfrom its database of stories to determine whatis an opportunityfor intervention.Its processingcan be dividedinto two phases: storage time, whennewstories are put into the system and the systemconsiders howit mighttell them;andretrieval time, whena studentinteracts with the GuSSsystem and and SPIEL watches for opportunities to give tutorial feedback. Figure 1 shows these phases. At storagetime: 1. Indicesare attachedto eachstory in the database. 2. SPIEL’sstorytelling strategies are applied to each index.If the strategyis applicableto the index,a set of rules is generated that will recognize an opportunityto tell that story using the storytelling strategy. Atretrieval time: 1. Opportunity-recognition rules are matchedagainst the state of the simulation. 2. Whenthe rules for a particular combinationof story andstrategymatchsuccessfully,the story is retrieved. 120 . After retrieval, natural languagebridgeand codaare generatedto integrate the story into the student’s current context. Indices for educational stories SPIEL’sindices are created manually.Sincethe stories are in video form, automaticprocessing wouldentail speech (and possibly gesture) recognition as well as natural languageunderstanding.SPIEL’sdesigntherefore calls for a humanindexerto watcheach story being told and use an indexing tool to compose indices that capture interpretationsof the story’s meaning. Theseinterpretations havethe general form, "Xbelieved Y, but actually Z," which is a form of anomaly. An anomaly is a failure of expectation that requires explanation (Schank 1982). Typically, anomalous occurrencesare whatmakestories interesting and useful, and they are a natural wayto summarizewhat a story is about. Anomalies are especially importantin stories about social activity since students are learning what expectations they should have and how to address expectationfailures. The anomaly forms the core of SPIEL’s indexing representation. Considerthe story transcribed here whose indexis shownin Figure2. I wentto this auto glass place onetime whereI had the biggest surprise. I walkedin; it wasa big, burly man; he talked about auto glass. So we were workingon a display ad for him. It waskind of a rinky-dinkshop and there wasa TV playing and a lady there watchingthe TV.It was a soap operain the afternoon.I talked to the mana lot but yet the womanseemedto be listening, she was Setting: step-of-sales-process(salesperson, pre-call) business-type(service) business-size(small) business-partnership(auto-glass man,wife) marded(auto-~lass man,wife) sales-target(salesperson,autc-~lass company) Anomaly:l salesperson [ assumed ii MeI theme I Assumed ~ ’ Actual Theme: housewife business panner Goal: hospitalit~ help in decision Plan: small talk evaluatepresentation Result: positive evaluation Pos.side-effect: sale for salesperson Neg.side-effect: Fit, ure 2. Indexfor "WifewatchingTV"story_ asking a couple of questions. She talked about the soapoperaa little bit andaboutthe weather. It turns out that after he and I workedon the ad, he gaveit to her to approve.It turns out that after I broughtit backto approve,she approvedthe actual dollar amount.Hewas there to tell me about the business, but his wife wasthere to hand over the check. So if I hadignoredher or had not givenher the time of day or the respect that she was deserved, I wouldn’thave madethat sale. It’s important when youwalkin, to really listen to everyoneandto really payattention to whateveris goingon that yousee. Theindex contains ¯ the anomalyin the story, whichcan be phrased as "the salesperson assumedthe wife wouldhave the role of housewife,but actually she was a business partner." ¯ the setting, the story’s position within the overall social task, includinga representationof the social relationshipsbetweenthe actors in the story, and ¯ intentional chains surrounding and explaining the anomalousoccurrences. SPIEL’sindices are considerably more complexthan those typically found in case-based reasoning systems. Thefact that SPIEL’scases are video clips is part of the reason. Its indices have to say everything that SPIEL needsto knowaboutits stories since the stories themselves cannot be evaluated. Anotherreason that SPIEL’sindices showcomplexityis the task of educational storytelling itself. SPIELhas a variety of reasonsfor telling stories, each of whichrequires a slightly different perspectiveon the index. SPIELneeds complexindices to meet the demands of a variety of strategies. Thethree strategies I describe next showsomeof these uses. Theothers can be foundin (Burkein preparation). 121 The Warn about plan storytelling strategy Astorytelling strategy is a wayof using a story to teach. Consider Example1, which shows the Warnabout plan strategy in action. Thestudent is pressing for a large ad campaign,muchlarger than the client needsor can really afford. The customeris an evasive type, and does not immediately reject the idea: instead, he stalls. Thestudent couldlose a lot of time in a futile effort to makethis sale. At this point, the storyteller interveneswith a story about an analogoussituation where, instead of stalling, the customer rejected the ad program and rebuked the salesperson. Telling the story at this point helps the student identify the problembefore goingtoo far along in this direction. Anaccurate assessmentof howfar to let the student go wouldrequire a great deal of knowledge about the scenario the student is engagedin, moreknowledge than is available to SPIEL.GuSS’sopen-endedsimulation design does not allow any quick read-out of what outcomesare likely and howfar awaythe student is from obtaining them. SPIEL would have to perform a complete envisionment of possible simulationoutcomesto find out, moreprocessing than can be accomplished in the midst of student interaction. Instead, SPIELuses prototypical knowledge about inconclusive outcomes. For SPIELto tell the "Wouldyou buy this ad?" story in the example,it must knowthat overcoming the customer’sstall in this kind of case is time-consuming andnot particularly educational. At storagetime, eachstorytelling strategyselects stories that are compatiblewith its educationalrole. Astory with a good outcomecould not be sensibly used with Warn about plan, for example. The second task of the storytelling strategy is to generate, for each compatible story, a Recognition Condition Description (RCD), representationthat describesa situation in whichthe story ¯ Thestudent doesn’t gather very muchinformationat the pre-call stage. ¯ Backat the office, the studentpreparesa very large ad campaign. ¯ Whenthe student presents this to the client, the client says"Youknow,I really haveto talk to Edaboutthis..." and is inwardly very doubtful of the value of such a large expenditure. Storyteller: Astory abouta failure in a situation similar to yours... Youmadea recommendationthat was muchlarger than Dave’sexpectations. Here is a story in whichdoing that led to problems: "I remember myfirst year 1970.I wason myfirst Yellow Pagesales canvass. In those days, they didn’t give youa lot of time to showability andI wasn’tdoingvery well. Mymanagertold methat I hadone weekto start producing or they weregoingto let mego. "I called on a graphicartist in Indianapolis.Hehada 2HS, a one-inchad. I walkedin, askedtwominorquestions, and I laid downa quarter-pagepiece of spec in front of this manand told him he needed this ad. The manlooked at meandhe said, ’Would youbuythis ad?’ Heturnedit right back to me. I didn’t knowwhat to say. It shockedme. Finally, I said, ’No, I wouldn’t.’Hedidn’t needa quarter page; a one-inch ad is what he needed. The gentleman proceededto give it to me,up oneside anddownthe other. Hetold methat I was there for myowngreedyinterests, trying to makecommissions instead of taking care of his advertisingand caring abouthim. Hesaid, ’If you’reever goingto makeit in this business,you’dbetter start paying attention to yourcustomers.’ "I walkedout of that call a different salesmanbecauseI realized then that the only wayto sell YellowPagesis to sell whatthe customerneeds, not what I need. Learning that lesson tamedmysales career around.I started making sales, andby the end of the canvass,I wasoneof the top producers." Youmadea recommendation that is muchlarger than the client’s expectations.Thatmightnot be a goodidea. Example1. Telling "Wouldyou buy this ad?" using the Warn aboutnlan strategy could be told using the strategy, a tutorial opportunity providedby the story. Computingthe RCDinvolves makinginferences based on the story’s index, essentially askingthe question"What wouldhave to happenfor this story to be goodto tell?" The answer to this question is different for every storytelling strategy. For Warnabout plan, the situation shouldbe pretty muchthe sameas in the story: the student should have overshot the markby designing a large ad program.The difference is that the student should be unlikelyto discoverthis error. Immediate rejection, which is the bad outcomeof the story, is not what should be looked for. Instead, the system needs an inconclusive outcome that bodesill. Whatoccurs in Example1 is one kind of outcomeof this kind: if the personthe studentis talkingto has authorityto 122 buy and that person defers the decision to someoneelse, this is usually an evasive maneuver,not motivatedby a real needfor consultation.Also, if the customerdelaysthe decision:"I don’thavelime for this now.Let’s talk aboutit next week."Neitherof these objectionsare associatedwith any of the importanteducational goals of the system, so interventionis justified. A descriptionof the RCD that characterizes the tutorial opportunitywouldthereforelook somethinglike this: WHEN the student is closing the sale andspeakingto someonewhois the decision maker, LOOK FORthe student to present a very large ad program, the client to havea negativebelief about that program,and the client to defer the decisionto another, or the client to put the decisionoff to anothertime, THENTELL"Wouldyou buy this ad?" ASa "Warnabout plan" story. This is the formatof an RCD.It contains a trigger (the "when"part) that describesthe conditionsunderwhichthe opportunity becomespossible, usually a function of the stage of the sales processthat the studentis in, combined with characteristics of the other agents involvedin that stage. If the triggeringconditionsare met,the systemtries to gather evidencethat similar intentionsare at workin the simulation. Warnabout plan can also be used to showeffects that wouldnot appear within the scenario. For example,scare tactics maypersuadea customerto buy once, but they hurt the client relationship and eventually result in lost business. This problemdoes not showup in YELLO since a scenario ends whenthe student makes one sale. A student whosuccessfullyuses this tactic in the simulation might think it is a good idea in general unless the storyteller can showan exampleto demonstrateotherwise. Another use of the storytelling strategy is in the generationof the natural languagetexts that surroundand explain the story. As shownin Example 1, there are three parts to the explanation:the headline, for the attentiongetting initial statement; the bridge, the introductory paragraphexplaining whythe story has comeup; and the coda, that closes the story presentation with a recommendation or evaluation. Eachstrategy has a set of natural languagetemplatesfor these texts, whichare filled in by generatingappropriatephrasesat retrieval time. The Demonstrate risks storytelling strategy There is tremendousvariability in the social world. Approachesthat have always workedmayfail in a new situation for no apparentreason; unlikely strategies may fortuitously succeed.Bothin GuSS’ssimulator andin real life, the learner gets to see only one outcomeat a time. Oneimportantrole that real-world stories can play is to illustrate the range of real experience by presenting counter-examplesthat contrast with what is happeningto the student. Studies of apprenticeshipsituations indicate that experts often use stories in exactly this way(Lave Wenger1991). TheDemonstraterisks strategy showsthat a successful result in the simulationis not alwaysrepeatablein the real world. Considerthe situation wherethe student closes a sale, but continuesconversingwith the customer.This is risky, since it gives the customer an opportunity to reconsider. To show the risk, SPIEL tells about a salespersonwholost a sale througha careless remarkmade after a sale wasclosed. This strategy resembles Warnabout plan because it uses a story about a failure to warn the student. The difference is that the Warn aboutplan strategy is used whenthe simulation is not likely to give immediate feedback about the student’s actions. The Demonstrate risks strategy looksfor the simulationgivingfeedback,but of the wrongkind. It waits for the simulation to do somethingopposite from the outcomefound in the story, and tells a story to show howthe real world can be different fromthe simulation. This approachto retrieving counter-examplesdiffers from existing CBRsystemsthat use contrasting examples, such as HYPO (Ashley & Rissland 1987). HYPO retrieves cases based on similarity along certain dimensionsand buildsa structure, the claimlattice, of the retrievedcases, enablingit to identify contrasts along other dimensions. SPIELuses its knowledgeof contrasts in the retrieval processitself, so that it onlyretrievesthosecasesthat make the needededucationalpoint. To arrive at the recognition condition description for such a tutorial opportunity,SPIELmustlook for a result that wouldbe opposite fromwhatoccursin the story. The story described aboveshoweda salespersonlosing a sale during conversationafter the close of a sale. SPIELuses an opposite-finding inference mechanismto identify a outcomewhich is opposite from the one shownin the story. Thestory is a goodcounter-example if the student doesa similar action, but doesnot lose anything. The "Warn about assumption" strategy Newcomersto a social domain may inappropriately transfer expectationsfromthe rest of their social lives. A newsalespersonmaythink, for example,that a friendly, talkative customeris morelikely to buythan a "get down Storytellingstrategy Summary Tell a story aboutan unsuccessfulplan Warnabout plan whenthe student has begunexecutinga similar plan. Demonstraterisks Warnabout assumption to business"type, whenin manycases, the oppositeis true. Sincethe indices usedfor stories incorporatethe difference betweena person’s view of the world and howthe world actually turned out to be, SPIELis well poised to help studentsby pointingout their unrealisticexpectations. If the programhas evidence that the student has a particular assumption, the Warnabout assumption strategy calls for it to tell a story about a time whena similar assumptionwas wrong.Supposethe student has an opportunityto gather informationabout a client’s business fromthe client’s spouse,but doesnot take that opportunity. It is reasonableto infer that the student assumesthat the spouse does not have an importantrole in the business. SPIELcan tell the "WifewatchingTV"story at this point, showinga case wherea similar assumptionprovedwrong. I call this type of strategy is a perspective-oriented strategy, becausewhat is important is the contrasting perspective in a story. Recognizingstories that are relevant in this wayis moredifficult than recognizing stories that contain similar actions. Perspective-oriented strategies need knowledgeabout howcertain actions are characteristic indicators of the student’s mentalstate. Students do not alwaysact in waysthat clearly indicate their beliefs, but if they do, the systemshouldbe prepared to respond. In this story, the salespersonassumes,in an information-gathering context, that the client’s spousewill not havea business role. Whatactions wouldindicate such an assumption on the student’s part? In information gathering, the primarytask is to ask questions about the client’s business.If the studentfails to ask suchquestions of someonewhengiven the opportunity, this is a good indicationof an assumption on the student’spart that there is no information to gather. Anotherpossible indicator wouldbe a deliberate act on the student’s part to exclude the spousefromthe sales presentation. Conclusion Thethree storytelling strategies describedin this paperand summarized in Table 2 showsomeof the demandsthat the task of educationalstorytelling places on a case retrieval system. The retriever must not only respond to crucial similarities between a story andthe situation in whichit is told, but also differences, such as opposite outcomesas called for by the Demonstrate risks strategy, and Tutorial opportunity Lookfor similar setting, similar goal andsimilar plan. Lookfor a negative, but inconclusive outcome. The negative Lookfor similar setting, similar Tell a story abouta negativeresult of a outcomeof a plan. goal andsimilar plan. particular plan whenthe studenthas just Lookfor an opposite outcome. executeda similar plan but hadsuccess. Tell a story aboutan erroneousassumption An assumptionthat Lookfor similar setting. didn’t hold. Lookfor actions that are that someonemadewhenthe student appears to have madethe sameassumption. indicative of the assumption. Table2. Summary of the three storytelling strategies 123 Storyis about: The negative outcomeof a plan. evidential relations, used by Warnaboutassumption and other perspective-orientedstrategies that look for actions that indicatebeliefs. Thedesign of SPIELis a responseto these demands.It uses structured, strategic comparisons betweenstories and the situations in whichthey are told. Theseretrieval strategies are an explicit counterpartto whatis implicit in the similarity metrics used in standard problem-solving case retrieval: a theory of whatcases are useful for the task. Extensionsof the standardproblem-solving paradigm (Kolodner, 1989) and newapplications for case-based reasoning technology, such as education, entail new notions of utility, and with them,newmetrics that combine similarity judgmentsandother measuresof fimess. Putting such strategies to workneed not render case retrieval prohibitively inefficient. Since strategies are implementedas proceduresthat operate at storage time, SPIELperformsa minimum of inference at retrieval time, yet still remainssensitive to the strategic considerations involvedin recognizingtutorial opportunities. Acknowledgments The GuSSproject has prospered under the guidance of RogerSchank,Alex Kass and Eli Blevis. I wouldlike to acknowledge their contributionto the ideas in this paper. GuSSis the product of manymindsand efforts including those of Mark Ashba, Greg Downey,Debra Jenkins, SmadarKedar, Jarret Knyal, MariaLeone,Charles Lewis, John Miller, Thomas Murray, Michelle Saunders, JeannemarieSierant, and MaryWilliamson.Thesales and sales training organizationsof AmeritechPublishinghave also contributedto this researchin manyways. References Ashley,K., and Rissland, E. 1987. Compare and Conlrast, ATest of Expertise. In Proceedingsof the 6th National Conference on Artificial Intelligence. AAAI. Birnbaum,L. A. 1986. Integrated Processing in Planning and Understanding. PhD thesis, Dept. of Computer Science,YaleUniversity. Blevis, E. andKass, A. 1991.Teachingby Meansof Social Simulation.In Proceedingsof the International Conference on the LearningSciences. AACE. Burke, R. 1992. Knowledge acquisition and education: a case for stories. In Proceedings of the Symposium on Cognitive Aspectsof Knowledge Acquisition. AAAI. Burke, R. in preparation. Representation, Storage and Retrieval of Stories in a GuidedSocial Simulation. PhD thesis, Dept. of EECS, Northwestern University. Burke,R. and Kass,A. 1992.Integrating CasePresentation with Simulation-basedLearning-by-doing.In Proceedings of 14th Annual Conference of the Cognitive Science Society, 629-634.Hillsdale, NJ: LawrenceErlbaumAssoc. Domeshek,E. 1992. Do the Right Thing: A Component Theoryfor IndexingStories as Social Advice.PhDthesis, Dept. of ComputerScience, Yale University. Issued as 124 Technical Report No. 26, Institute for the Learning Sciences. Edelson, D. 1993 Learning from Stories: Indexing, Remindingand Questioning in a Case-based Teaching System. PhD thesis, Dept. of EECS, Northwestern University. Gentner, D. and Forbus, K. 1991. MAC/FAC: A Modelof Similarity-BasedRetrieval. In Proceedingsof 13th Annual Conferenceof the Cognitive Science Society. Hillsdale, NJ: LawrenceEflbaumAssoc. Hammond,K. 1989. Case-based Planning: Viewing Planningas a Memory Task. Boston: AcademicPress. Kass, A. and Blevis, E. 1991. Learning Through Experience: An Intelligent Learning-by-DoingEnvironment for Business Consultants. In Proceedings of the Conference on Intelligent Computer-AidedTraining. NASA. Kass, A.; Burke, R.; Blevis, E.; and Williamson,M. 1993. Constructing Learning Environmentsfor ComplexSocial Skills. Journalof the LearningSciences.Toappear. Kolodner,J. 1989. JudgingWhichis the ’Besf Casefor a Case-basedReasoner. In Proceedings of the Case-Based Reasoning Workshop. DARPA/ISTO. Kolodner,J. and Jona M. 1992. Case-basedReasoning.In Encyclopedia of Artificial Intelligence(2nded.), S. Shaprio (ed.). NewYork: Wiley. Lave, J. and Wenger, E. 1991. Situated Learning: legitimate peripheral participation. NewYork:Cambridge UniversityPress. Owens,C. 1989. Integrating Feature Extraction and MemorySearch. In Proceedings of the 11th Annual Conferenceof the Cognitive Science Society. Hillsdale, NJ: LawrenceErlbaumAssoc. Schank, R. C. 1982. Dynamic Memory: A Theory of Learning in Computersand People. NewYork: Cambridge UniversityPress. Schank, R. C. 1991. Tell Mea Story. NewYork: Charles Scribner’sSons. Stevens, S. 1989.Intelligent Interactive VideoSimulation of a CodeInspection. Communications of the ACM,32(7), 832-843. Thagard, P.; Holyoak,K.; Nelson, G.; and Gochfeld,D. 1990.Analogretrieval by constraintsatisfaction. Artificial Intelligence, 46(3): 259-310. Waltz, D. L. 1989. Is Indexing Usedfor Retrieval7 In Proceedings of the Case-Based Reasoning Workshop, DARPA/ISTO. Williams,S. M. 1993.Putting Case-Based Instruction into Context: Examplesfrom Legal, Business, and Medical Education.Journalof the LearningSciences. To appear.