The Role of Perspective in Responding to Property Misconceptions

1. Introduction

Aa a uaer iDteraeta with a databaee or an expert eyetem, e/he may attribute a property or property value to an objeet that that objeet doea not have. For iDatance, imaaine the foUowiDa query to a databaee.

1. Give me the IRJLL-NO of all DESTROYERS whoee MAST-HEIGHT ie above 190.

Ir a system were to evaluate such a query, it miaht rmd that there are no such ships in the databaee. The reason for thia ia that the user has queried a value of the property MAST-HEIGHT that it cannot have for the objeet DESTROYER. I term thia enor a propertv mi1conception. Upon encounterina such a query, even a very cooperative system could only reapond:

1. There are no DESTROYERS iD the databaee with a MAST-HEIGHT above 190. Would you like to try aaain!

In most caaea, however, thie ie not th~ way a human would respond. A study or human/human transcripta reveala that a human convenational partner often triea to aet at the cause of the misconception and otrer additional information to correet the wrona information. The additional information often takea the form of a correet query that is a possible alternative to the user's query. In thie paper I describe some or the knowledae and reasonin& that are neeeeaary for a naturallanauaae interfaee to a databaee or expert system to mimic thie human behavior.

In the above query, since there ie an objeet similar to a DESTROYER that has the value of IRJLL-NO aiven, the user'a misconception may reault rrom his/her confusiDa the two objeeta. Hence a reasonable reaponse would be:

1. All DESTROYERS iD the databaee have a MAST· HEIGHT between 85 and 90. Were you thiDkiD& or an AIRCRAFT-CARRIER!

Notice the atrate&Y ueed to correet the miaeonception ia to ( 1) deny (implieitly) the property /value aiven, (2) aive the correapondiD& correet iDformation, (3) augeat an alternative query contaiDiD& the objeet the user may have confueed with the miaeonception objeet.

. In other aituatione, a reasonable alternative query mi&ht mvolve the aame objeet the user asked about, with a different property/value pair. Thia ia the caee iD the foUowiDa query.

0 Thio work il pariially tupporied b;y lhe NSF cranl *MCSSJ-07200 and b;y lhe ARO rranl DAA2G-84-K-004J .

1. Give me the IRJLL-NO of aU DESTROYERS whoee MAST-HEIGHT ia above 3500.
2. AU DESTROYERS iD the databa:~e have a MAST· HEIGHT between 85 and 90. Were you thiDkiD& of the DISPLACEM!i:NT!

Thie reaponse ia similar to the one aiven above excl!pt that the alternative query auueata an attribute rather than an objeet whieh may have been confueed.

In aeneral, there can be two major reasone why a wrona attribution may occur. Either (1) the user hae the wrona objeet­ that ie, e/he hae confueed the objeet beina diaeuaeed with a eimilar objeet or hae reasoned (falsely) by analo&Y from a eimilar objeet; or (2) the user hae the wrona attribute - that ie, a/he hae confueed the attribute bein& diaeuaeed with a eimilar attribute. U one: of. theee two can be eeen ae likely iD a aiven aituation, then a re:viaed query can be augeated which mentions the aimilar objeet or the similar attribute.

To propose alternative queriea, a syatem muat have a method ror determiDiD& eimilarity or objeeta and attributee. In thia paper I will rocus on reaponsea iDvolvina objeet confusion; thua I will examine a similarity metrie for objeeta. In the next aeetion auch a similarity metrie ie introduced. The foUowiDa aeetion introducea a new notion of objeet penpeetive which ia needed to provide the similarity metrie with some neeeeaary contextual iDformation, iD particular, attribute salience ratinaa. FiDally, an example or how perspeetive information and the aimilarity metrie can be ueed to aive reasonable responsea to miaeonceptione iDvolviD& objeet propertiea ia aiven.

2. Objeet Similarity

As was shown above, in order to reapond etreetively to property miaeonceptions, we must have a method for determinina object similarity. Objeet similarity has previously been shown to be important in tasb such ae organizina explanationa 161, otreriD& cooperative reaponsea to praamatically ill-formed queriea 121, and identifyina metaphors 101. In the above systema the similarity or two objeeta ia baeed on the dietance between the objeeta iD the generalization hierarchy. One problern with thia approaeh ia that it ie conte:d in11oriont. ·· That ia, there ia no way for contextual information to atreet aimilarity judamenta.

However, Tvenky 181 propoaea a meaeure or objeet aimilarity baeed on common and diejoint reaturea/propertiea of the objeeta involved, whieh enablea contextual information to be taten into aecount. Tveraky'aaimilarity ratina for two objeeta a and b, where A ia the set or propertiea aeaoeiated with objeet a and B ia the set or propertiea aeaoeiated with objeet b, can be expreaeed ae:

s(a,b) = 6f(A n B) - af(A - B) - ßf(B - A) for some 6, a, ß )- 0. Thia equation aetually dermea a family or similarity scalea where 6, a, and ß are parametera which alter the importance or eaeh pieee or the equation, and r maps over the reaturea and yielda a aalience ratin& ror eaeh. The equation statea that the similarity or two objects is 80me function or their common features minus 80me function of their disjoint features. The importance or each reature involved (determined by the runction ,, and the importance or each piece or the equation (determined by '· a, and ß) may change with context.

00 See 161 ror addilional probleiDJ and dilcunion or lhil poinl.

Previous worlt [4, 71 has discussed the errect or ·rocus· on the prominence of objects. Focusing algorithms can be adapted to set the values of 8, a, and ß. For instance, iC object a is "in focus· and object b is not, then the features of a should be weighted more heavily than the features or b. Thus we should choose a > ß 80 that the similarity is reduced more by features of a that are not shared by b than vice versa.

The problern then is to determine r. Other worlt (3, 91 has band encoded salience values for the attributes of individual objects in the ltnowledge base, errectively setting the r function once and for all. This approach, however, is not surricient since salience values must change with context. The following examples in which two objects (Treasury Billsand Money Marltet Certificates) are compared in two different circumstances, illustrate the importance of context on the similarity rating.

Consider 80meone calling an expert rmancial advisor to see iC she can better invest her money. She begins by telling the expert where her money is:

1. We have $40,000 in money marltet certificates. One is coming due next weelt for $10,000... I was wondering iC you thinlt this is a good savings ...
2. Weil, l'd lilte to see you hold that $10,000 coming due in a money marltet rund and then get into a Ionger term money marltet certificate.
3. Hm ... weil I was just wondering, what about a treasury bill instead!
4. That's notabad idea but it doesn't replace your money marltet certificate in any way - it's an exact duplicate. They're almost identical types of instruments- 80 one, as rar as I'm concerned, is about the same as another.

Now consider how the same two objects can be seen quite differently when viewed in a different way. Imagine the following conversation:

1. I am interested in buyiog some US Government Securities. Now I was thinlting of Money Marltet Certificates since they are the same as Treasure Bills.
2. But they 're not - they are two very dirrerent things. A Treasury Bill is backed by the U.S. Government: you have to get it from the federal reserve. A Money Marltet Certiricate, on the other band, is baclted by the individual baolt that issues it. So, one is a Government Security while the other is not.

In the first example the objects are viewed as savings instruments. This view highlights attributes such as interest-rates and maturity-dates that are common to Treasury Bills and Money Marltet Certificates. This highlighting causes the two instruments tobe seen as "identical· . In contrast, the second example views the objectli as instruments issued by a particular company or organization. In this case attributes such as issuing-company and purchase-place are highlighted. Since these highlighted attributes are dirrerent for the two objects, the objectli are seen as being quite different.

As the examples illustrate, a context-free metric of similarity is not surricient; contextual information is needed. A notion of object perspective, introduced below, can capture the needed contextual ioformation. In particular, perspective accountli for how the r function (the assignment or salience values to various attributes) changes with context.

3. Pers.,ective

(4, 1) note that the same object may be viewed from different perspectives. For instance a particular building may be viewed as an architectural worlt, a home, a thing made with bricb, etc. According to this worlt, an object viewed from a particular ~rspec~ive is seen as having one particular superordinate, although m fact at may have many superordinates. The object inherits properties only from the superordinate in perspective. Therefore different perspectives on the same object cause different properties to be highlighted.

Although this notion or perspective is intuitively appealing, in practice its use is rather difficult since it hing es on the . use of a limited inheritance mechanism. The problern is that attributes may be inherited from the top of the generalization hierarchy, not just from immediate superordinates. So, an object's perspective involves not just one superordinate but a chain of superordinates. Therefore one must not only determine what perspective a particular object is being viewed from, but also what perspective its superordinate is viewed from, and 80 on. As one continues up the hierarchy in this fashion, the definition or perspective as viewing an object as a member of a particular superordinate becomes less and less appealing.

In addition, this notion of object perspective says nothing about the density of the generalization hierarchy. That is, in 80me situations the immediate superol'dinate of an object (and the properties it contributes) may be ignored. For example, even though a whale is a cetacean (a dass of aquatic mammals induding whales and porpoises), this dassification (and all attributes contributed by the dassification) may be ignored in 80me situations in which the important attributes instead are inherited from a superordinate of cetacean, say, mammal. In other situations, the dass ·cetacean· may be central. The notion of object perspective outlined above has no way of determining whether or not certain superordinates should be ignored or induded.

Here I introduce a new notion of perspective which is able. to handle both the assignment of dirrering salience values and the density problem. In this notion, perspectives sit orthogonal to the generalization hierarchy. Each comprises a set of properties and their salience values. A number of perspectives must be defined a priori for the objects in a particular domain. The specification of perspectives, just lilte the specification or an object taxonomy' must be done by a domain expert. Knowledge of useful perspectives in a domain then, is part or the domain expertise.

With this new notion of perspective, when an object is viewed tbrough a particular perspective, the perspective essentially acts as a filter on the properties which that object inherits from its superordinates. That is, properties are inherited with the salience values given by the perspective. Thus properties of the object which are given a high salience rating by the perspective will be bigbligbted, while tbose whicb are given a low salience value or do not appear in the perspective will be suppressed. The density problern is handled by ignoring those Superordinate concepts which contribute only attributes suppressed by the current perspective.

4. Using Perspective to Determine Responses

. Perspective information can be used with Tverslty's similarity metrac to help determine alternative queries to a query containing a misconception. To see how this worlts, consider a domain containing the following three objectli with the attributes shown:

Money Marltet Certificates Maturity: 3 months Denominations: $1,000 Issuer: Commercial Bank Penalty for Early Withdrawal: 10% Purehase Place: Commercial Bank Safetv: Medium Trell8ury Bills Maturity: 3 months Denominations: $1,000 Ieeuer: US Government Purehase Plaee: Federal Reserve Salety: Hi&h

Tre118uQ' Bond

Matur1ty: 7 yean Denominations: $500 Ieeuer: US Government Penalty for Early Withdrawal: 20% Purehase Plaee: Federal Reserve Salety: Hi&h

and the foUowing perepec:tivee:

Savinss Instruments Maturity - high Denominations - high Safety - medium

leeums Company Ieeuer - high Salety - high Purehase Plaee - medium

Notice that the perepec:tive of Savings Instrumente highlighte Maturity and Denominations, and 110mewhat highlighte Salety. This indicatee that when people are diseuseins securitiee 118 savings instrumente, they are moet intereeted in how long their money will be tied up andin what denominations they can save their money. The perepec:tive or lseuing Company. on the other band, highlighte dirrerent attributee. When securitiee are discuseed from thia perspec:tive, things like wbo the iasuer or the security ia and how safe a security iasued from that company ia, beeome important.

Suppoee the perepec:tive ia Savings Instrumente and the user says:

U. What ia the penalty for early withdrawal on a Trell8ury Bill!

Thia query indicatee that tbe user bll8 a misconception since s/be has attributed a property to Trell8ury Bills that they do not bave. One reasonable correction to the query would contain an alternative query which to replaeee Trell8ury Bills with another object tbat bll8 the property spec:ified and ia similar to Trell8ury Bills. Tbe system may reason tbat both Money Marltet Certificatee and Treasury Bonds have the penalty spec:ifJed, and 110 check to see ir either or these objecte ia similar to Trell8ury Bills. Notice that the Savings Instrumente perepec:tive highlighte attributee common to Treasury Bills and Money Market Certificatee (they have the same Maturity and Denominations), 118 weil 118 attributee diajoint to Trell8ury Bills and Trell8ury Bonds (they have dirrerent Maturity and Denominations). Using these salience valuee, the similarity metric will find that Money Martet Certificatee are very similar to Trell8ury Bills while Trell8ury Bonds are very dirrerent. Thus Money Market Certificatee will be deemed a probable object or confusion and the (ollowing correction may be orrered:

1. Trell8ury Bills do not have a penalty for early withdrawal. Were you tbinting of a Money Martet Certiftcate?

Notice that ir the perepec:tive bad instead been leeuin& Company, which higblighte attributee common to Treasury Bills and Treasury Bonds and diajoint to Trell8ury Bills and Mont)' Market Certiftcatee, the moet reasonable reeponse would be:

1. Treasury Bills do not have a penalty for early withdrawal. Were you thinting of a Trell8ury Bond?

Selecting the appropriate perepec:tive ia in itself a dirficult question which ia currently under investigation and will be reported in 151. Certainly important in tbe selection proeedure will be tbe attributee that bave entered into the convereation 110 rar: theee attributee should be or fairly high salience in the selected perepec:tive. Other cluee to the selection proeese include tbe objecte under discuseion, tbe superordinatee which contribute the attributee under discuseion to these objecte, and the current goals or tbe user.

6. Coneluslon

In thia paper we have seen that a context-dependent similarity metric ia needed in order to respond adequately to miaconceptions involving the properties of an object. Such a metric has been suggested and a notion or perepec:tive bll8 been introduced to account for 110me or the contextual information required by the metric. These notions have been shown to account for dirrerencee in the way a particular misconception ia best corrected in two dirrerent circumstancee.

8. Aeknowledgemente

I would lite to thant Julia Hirschberg, Aravind Joshi, Martha Pollack, Etbel Schuster, and Bonnie Webher ror their many commente and discuseions concerning tbe direction of thia research and tbe content and style or tbia paper.

7. Referenees

111 Bobrow, D. G. and Winograd, T. · An Overview of KRL, a Knowledge Representation Language. · Cognitive Science 1, 1 (January 1077), 3-46.

• 121 Carberry, Sandra M. Understandins Pragmatically nJ..Formed Input. 10th International Conference on Computational Linguiatics & 22nd Annual Meeting of the Aseoeiation of Computational Linguiatics, Coling84, Stanford Univereity, Ca., July, 1084, pp. 200-206.
• 131 CarbonneU, Jaime R. & Collins, AUan M. Mixed-Initiative Systems For Training and Deciaion-Aid Applications. Tech. Rept. ESD-TR-70-373, Electronics Systems Division, Laurence G. Hanscom Field, US Air Force, Bedford, Ma., November, 1070.
• 141 Grosz, B. Foeusing and Description in Natural Language Dialoguee. In Element' of Di1couru Underltanding, A. Joehi, B. Webher & I. Sag, Ed.,Cambridge Univereity Prese, Cambridge, England, 1081, pp. 85-105.
• 151 McCoy, K.F. Correcting Object-Related Misconceptions. 1085. Forthcoming Univereity of Pennsylvania doetoral theeia
• 161 McKeown, K. . Generating Natural Language Tezt in Re1pcm1e to Que1ticm1 About Databa1e Structure. Ph.D. Tb., Univereity of Pennsylvania, May 1082.
• 171 Sidner, C. L. Foeusing in the Comprehension of Defmite Anaphora. In Computaticmal Modell oJDi1cour1e, Michael Brady and Robert Berwict, Ed.,MIT Prese, Cambridge, Ma, 1083, pp. 267-330.
• 181 Tverety, A. ·Features of Similarity.· P111chological Review 8-4 (1077), 327-352.
• 101 Weiner, E. Judith. "A Knowledge Representation Approach to Understandins Metaphore.· Computaticmal Linguiltie~ 19, 1 (January · March 1084), 1-14.