Defining Salience in Case-Based Arguments

1 Introduction

A recurring theme of AT has been to design systems that can tailor descriptions of knowledge base objects to suit alternat ive points of view and contexts. In 1975, for ex­ ample, Minsky proposed frame representation sd~emes, in part, as a way of efficiently handling changes in view­ point, either an observer's physical perspective on a room or his metaphorical perspective in a non v.isual do­ main [Minsky 75]. By 1995, researchers on the "Knoe­ sphcre" project speculated that their comprehensive, in­ teractive encyclopedia would be able, at the reader's choice, to describe tl1e same basic reference information from a variety of styli7:ed viewpoints such as "l':rnpha­ sizeTheory", "NoMath," or "Teleology" [Lenat 82].

As this perer111ial theme iiJust.rates, one hallmark of in­ telligence is knowing what. aspects of an object. are salient from a particular viewpoint. Salience means slrikingness or emphasis. As viewpoints change, different features of an object stand out.

"Currently, the author is a Visiting Scientist at the IBl\1 Thomas J. VVatsnn Research Center, P.O. Box 21.8, York­ town Heights, New York ]0598. Formerly, the author was a postdoctoral researcher and graduate student at the De­ partment of Computer and Information Scien<:e, University of Massachusetts, Amherst, l\1assachusetts 01003. This work was suppor ted in part by: the Advanced Research Projects Agency of the Department nf Defense, monitored by the Of­ fi ce of Naval Research under contract no. NOOOH- 8·1-K · 0017; the University Research Initiative, award no. NOOOH86 -K- 0764; and an lBM Graduate Student Fellowship.

The goal has been to design systems in which the con­ text guides selecting which features to emphasize while, at the same lime, minimizing the computational com­ plexity of the selection process. Some progress has been made. Research in explanation has focused on slanting descriptive explanations to reflect the ques­ tioner's goals as evidenced by the question's perspec­ tive [f\1cKeown 85]. Ordering objects in a scene accord­ ing to "visual" salience has been explored as a guide to planning natural language descriptions of the scene [McDonald 82].

Determining salience is especially important in mak­ ing argu menl.s. Given the reality that most positions have weaknesses as well as strengths, a debater must know what facts help the argument or hurt it, what facts should be emphasized, mentioned or simply ig­ nored. Attorneys are notoriously adept at describing (some might say twisting) the same facts from oppos­ ing vantage points to support their clients' position. For example, a city comm issioner hires the daughter of the judge who presides over the commissioner's lover's di­ vorce. The prosecu lor calls it a bribe; defense counsel calls it a coincidence. Beside characterizing the facts differently, the attorneys must decide which facts to em­ phasi:w (e.g., did the daughter have relevant employment experience or had she never before been able to hold down a job.)

A number of AT researchers have approached the problem of salience in arguments by identifying var­ ious structures to elucidate s11pport and attack rela­ tions and guide inference, (e.g., argument graphs and "molecules" [Birnbaum 82], argument trees [Cohen 83], and conversational moves [ReichmanAdar 84].) In gen­ eral, the techniques have not been computationally pow­ erful enough to generate realistic domain arguments.

Case-based reasoning techniques allow salience deci­ sions to be made in the context of a more limited kind of argument , case-based arguments, arguments in which prior cases arc cited as justifications for deciding a prob­ lem situation. Salience has special significance in case­ based reasoning where cases are used either as models for problem solving, examples for explanation, or argu­ ment justifications. For instance, Kolodner and Simpson demonstrated how prior cases can focus a problem solver on those features of the problem situation that are salient for its solution [Kolodner 85].

Interestingly, the reverse is also true: when a past case is to be used in an argu mcnt about how to decide a prob­ lem, the features of the case that are salient depend on the particular facts of the problem situation, the view­ point of the arguer and the argument context. in which it is used.

A case-based reasoning program, HYPO, describes objects in a case knowledge base ("CKB") in a con­ textual way so that dilfcring views of a case are sup­ ported and reasoned with and decisions about salience are made without complex inference. HYPO is an adver­ sarial case-based reasoning program described in detail in [Ashley 87a]. It makes arguments how to decide a problem by citing relevantly similar past cases as j ustifi­ cations. In the past we have described the ways in wltich HYPO selects the most relevant cases and employs them in arguments [Ashley 87h,Ashley 88].

II ere we focus on the narrower issue of how ll Y PO makes determinations of salience, that is, how it deter­ In ines what to say about a case that it cites in an ar­ gument. Since IIYPO draws factual analogies between a cited case and the problem, one aspect of salience is determining what factual similarities (or dilferenccs) arc relevant to the analogy. A not her aspect of salience, how­ ever, is determining which of the relevant similarities and dilferences to emphasize in a given context and how to characterize them to convey tltat emphasis.

2 Views and Contexts

In HYPO, determining salience of features depends on the viewpoint and context of the description of the case.

HYPO supports two points of view in an argument correspondi11g to the two warring sides in a lawsuit: the plaintilf and the defendant. The plaintiff complains that the defendant's actions arc illegal and caused injuries to tl1e plaintilf for which it seeks relief from the court. HYPO deals with lawsuits involving trade secrets claims in which tltc plaintilf complains that the defendant has gained an unfair competitive advantage by obtaining il­ legally plaintiff's commercial secrets.

IIYPO describes cases in four basic contexts common to case-based arguments in law, political decision mak­ ing, evaluating property and ordinary discourse (Sec dis­ cussion in [Ashley 87a].) llerc we describe the contexts in terms of legal argu rnents:

1. Cited Case to make a point in a legal argument: HYPO makes an argument that a side should win the lawsuit described in the problem situation by citing a past case (the cited case) as an authority. In other words, HYPO adopts the viewpoint of, let us say, the plaintilf, argues that the plaintilf should win because the problem is similar to the cited case where a plaintilf won and emphasizes tltc relevant factual similarities.
2. Distinguished Case in response to a point in ale­ gal argument: HYPO contests the argument point by distinguishing the cited case. That is, IIYPO

f 1 = Secrets-Disclosed-Outsiders (Helps defendant ( 6) to extent that plaintiff has disclosed secrets to outsiders)

j2 := Outsider-Disclosures-Restricted (Helps plaintiff (1r) if all disclosures to outsiders are confidential)

f:l = Disclosure-in-Negotiations (Helps defendant (6) if plain tiff disclosed secrets to defendant in negotiations)

Figure 1: Three Sample Factors

adopts the view of the opponent, the defendant, and asserts that the dilferences between the problem and the cited case are so great that the defendant should win. In this context, HYPO emphasizes the rele­ vant factual differences between the problem and the cited case, which · we will refer to as the distin­ guished case.

1. Counterexample to a cited case: Another way that HYPO contests an argument point is by citing a past case as a counterexample to the cited case. That is, taking the view of the defendant, H'I'"PO cites as a counterexample a case won by a defendant. HYPO emphasizes the features of the counterexam­ ple that counteract the force of the plaintilf's argu­ ment based on the cited case. As described below, there arc four kinds of counterexamples.
2. Target Case that motivates a hypothetical modi­ fication: HYPO suggests how to strengthen a side's argument by suggesting l1ypothetical variations of the problem. Taking the view of either plaintilf or defendant, HYPO suggests how to bolster the argu­ ment by focusing on new features that would make the problem more similar to a past case (the target case) which could then he cited in support of that side.

The program computes the most salient features to cmpl1asize in describing a case from the CKB when it is used in any of the above contexts as a cited case, distinguished case, counterexample or target case. The features that are salient in any of these contexts can be described in succinct computational terms. In order to present those definitions, we define relevant similarities and introduce a bit of formalism.

3 Defining Relevant Similarities

Jn JIYPO, the features of a case that matter are defined in terms of factors. A factor is a general collection of facts that tends to favor or hurt the plaintilf's argument in a particular kind of lawsuit. Factors have magnitudes, that is, a particular case may be a more or less extreme example of a factor. Factors are represented in HYPO with dimensions [Ashley 87a]. Each dimension has a set of prerequisites used to test if the factor applies to a case and a range of values over which the magnitude of the factor in a particular case may vary. Figure l shows examples of three factors in the trade secrets domain.

Crown Industries

Outcome: Defendant won Factors: { 1 1 1 3 } Magnitudes: M(Jbrown) =disclosures to 7 outsiders Midland Ross Outcome: Defendant won Factors: { 1 1 } I'VIagnitudes: 111(Jl1 ;dlandR.o .. ) =disclosures to .100 outsiders Data General Outcome: Plaintiff won Factors: { f 1 1 2 } Magnitudes: M(JgataGeneral): disci~>StHes to 6000 _outsiders A-1(1 DataGeneral)- all disclosures restncted Crown Industries - Variant Outcome: Defendant won Factors: { 1'! 2 j3 } Magnitudes: llf(.fbrown-vM) =disclosures to 7 outsiders !lf(f2,own-var) = all disclosures restricted

Cases in the database arc represented in HYPO es­ sentially as historical collections of factors, each with a particular magnitude, to which some authoritative de­ cision maker (i.e., the judge) has assigned an outcome, either for the plaintiff or the defendant. Figure 2 shows examples of four cases. Each case is described abstractly in terms of an outcome, a set of applicable factors, and the magnitudes of certain of those factors. 1

Relevant similarities among cases are defined compu­ tationally in terms of sets of shared factors. The set of relevant similarities, 5c,,c, between two cases, CJ and c2, is the intersection of the sets of factors that apply to each case. Jn symbols:

Intuitively, the relevant similarities are reasons why the two cases should be decided the same way. Since the same combination of factors applied to each, the cases

1 The following nntatinn will make it possible to speak more succinctly ahnu t factors and cases. Let:

F+ ::::::: the set of all factors, /, such that f generally favors plaintiff.

F_::::::: {f If generally hurts plaintiff}.

Fe; ::::::: {f I f applies to t:i}, (i.e., the set of all

factors, f such that f applies to case t:;.)

For a given factor fk :

1\.f(f:;)::::::: the magnitude n[ factor fk in case t:;.

"M(f:;) > 111(!:;)" means that the magnitude o[ factor fk is greater for the plaintiff in case t:; than in case Cj.

"\" ::::::: set difference.

presented similar strengths and weaknesses from plain­ tiff's viewpoint.

Relevant differences are defined computationally in terms of factors not shared and differences in magni­ tudes of shared factors. Assume that the outcome of c2 was in favor of plaintiff. The set of relevant differ­ ences between c1 and c2 , De,,e,, is the union of three sets, the pro-defendant factors that apply only to c1 , the pro-plain tiff factors that apply only to c2 and the shared factors which favor plaintiff more strongly in c2 than in CJ. Tn symbols:

Intuitively, tlte relevant differences are reasons why the two cases should be decided differently. All of these rele­ vant dilrerences make c1 a weaker case for plaintiff than c 2 . As a result, c2 is a weaker justification in a legal argument that plaintiff should win in c1 .

4 Selecting Most Relevant Cases

ln lJYPO, a case is analogous or "on point" to a problem if it shares some relevant similarities to the problem, p, (i.e., 5p,c; f: 0 . ) One case, c;, is "more on point" than another case, Ck, relative to a problem if the set of relevant similarities between the problem and c; is a superset of the set of relevant similarities between the problem and Ck (i.e., 5p,c; :J 5p,q .) The set of on point cases that are most relevantly similar to the problem, that is, the set of "most on point" cases, M 0 P, is defined as the set of all cases, c;, such that for each c;, there is no case that is more on point than c;. To put that another way, !tf 0 P is the set of all cases, c;, such that c; is on point and for all other on point cases ck, either c; and q ha,·e no similarities in common with the problem situation or c; is as or more on point than q, or neither is more on point than the other.

fntuitivcly, M 0 P contains the cases that are candi­ dates for the best cases to cite for either side in an argu­ ment how to decide the problem. Of all the cases in the CKB, they are closest in that they share with the prob­ lem the greatest overlap of factual strengths and weak­ nesses as represented by factors. For a given problem sit1mtion, having identified the factors that apply to it, HYF'O computes AfOP using the usual set operations.

5 Defining Salience within Contexts

With these concepts, one may succintly describe the fea­ tures of a case that are salient when it is used in any of the four argument contexts. To illustrate how the salience expressions work, Figure 3 presents HYPO's de­ scriptions of a single case, the Data General case of Fig­ ure 2, used in all four contexts, as a cited case, distin­ guished case, counterexample and target.

Cited Case: The salient features of a cited case (cc) are its relevant similarities to the problem situa­ tiou . More specificalJy, the set of relevant similarities

L. As Case Cited by Plaintiff:

Where: Even though: Plaintiff disclosed its product in formation to outsiders, plain tiff should win a claim for trade secrets misappropriation.

Cite: Data General

2. As Case Distinguished by Defendant:

Data General is distinguishable because: ln Crown Industries, Plaintiff disclosed its product. infonna­ tion in negotiations with defendant. Not so in Data General. In Data General, plaintiff's disclosures to outsiders were restricted. Not so in Crown Indus­ tries.

3. As Counterexamples Cited by Plaintiff:

Trumping:

Data General is more on point [than l\1idland Ross] and held for plaintiff where it was also the case that: plaintiff's disclosures to outsiders were rest.ricterl.

Boundary:

Data General held for plaintiff even though in Data General plaintiff disclosed its product in­ formation to more outsiders than in Midland Ross.

4. As Target Case for Plaintiff:

Plaintiff's response would be strengthened if: plain­ tiff's disclosures to outsiders were restricted. Cf. Data General.

Figure 3: Descriptions of Data General Case 111 Four Argument Contexts

between cc and p is the intersection of the sets of factors that apply to each (i.e., Sp,cc = F~ n Ji~c . See equa­ tion 1.) Figure :J, No . . 1 shows how HYPO describes the Data General case as cited case (cc), where the prob­ lem situation (p) is that of the Crown Tndn.!trie.! case of Figure 2.

HYPO's description of Data General as a cited case in No. 1 emphasizes the fact.or shared with the prob­ lem {1 1 involving disclosures) and ignores the un­ shared factor (1 2 involving restrictions on disclosures.) HYPO computes the salient features using expression 1:

Intuitively, this makes sense. To call attention to f 2 at this point would not be tactically sound since f 2 is not a relevant similarity but a relevant difference (i.e., it is a pro-plaintiff factor that Data General does not share with the problem.) It is up to the defendartt to point out the differences (which HYPO does as we see below.)

Note that HYPO prefaces the description in No. with the word "Eventhough" in recognition of the fact that although f 1 is a relevant similarity, plaintiff's dis­ closures to outsiders arc not normally reasons for decid­ ing in plaintiff's favor. Had there been some relevant similarities that favored the plaintiff, for example, that the plain tiff had taken substantial precautions to secure its secrets, HYPO would have stated them first. That is, HYPO would say, "Where: plaintiff took substan­ tial security precautions, Eventhough: plaintiff disclosed its product information to outsiders, plaintiff should win , Distinguished Case: The salient features in distin­ guishing a case are the relevant differences between the problem and the distinguished case, Dp,dc· As Figure 3, No. 2 shows, when HYPO describes the Data General case as distinguished case (de), where the problem situa­ tion (p) again is the Crown case, it focuses on factors that Data General does not share with Crown (J2 involving restrictions on disclosures and f 3 involving disclosures in negotiation.) Making the substitutions in definition (2) of relevant differences, the three terms are:

{} (M(fbataGeneral) "f .M(Ji:rown) {6000 disclosures is worse, not better, for plaintiff than 7 disclosures.)

Taking the union of the terms yields the set of salient differences between the Cmwn case and Data General,

Dcrown,Da.taGeneral:

HYPO argues that these differences not only account for the outcome in Data General but also show that the Crown case need not have the same result.

Note that the analysis allows HYPO to select those dif­ ferences in factor magnitudes that are salient in a given context. Although the magnitudes of Data General and Crown along f 1 are different, HYPO sensibly does not call attention to this difference. The difference does not help, indeed it hurts, the defendant's argument. Crown is much stronger for plaintiff than Data General in terms of the Jmmbers of disclosures to outsiders.

'l'he analysis also allows HYPO to point out those fac­ tors that are salient in a context because they do not apply to a case. For example, in the context of distin­ guishing Data General, F' is significant because it docs not apply to Data General. The importance of this "non­ feature" becomes apparent only in light of the process of distinguishing Data Gcnr:ral from Crown.

Counterexample: IIYPO supports four kinds of counterexamples, each with a somewltat different defi­ nition of the features of a case that are salient when citing it as a counterexample. Each kind of counterex­ ample involves a case that is "cor1trary" to the cited case in the sense that it had the opposite outcome, focuses on different salient features, and can be used to disparage the impact of the salient features of the cited case. The four kinds of counterexamples and the associated salient features are discussed below:

(1) Trumping: a contrary case that ltas more (in the sense of a superset of) relevant similarities to the prob­ lem than t.he cited case. This kind of counterexample

"trumps" the cited case. It. has all the cited case's sim­ ilarities and then some. IIYPO has a method for com­ puting the set of trumping counterexamples, TCX ·,cc· that can be cited by a side, 8, in response to a case, cc, cited by s's opponent. TCX.,cc is the set of all most on point cases, c;, such that the outcome of c; favors sand c; is more on point than cc.

The salient features, the ones that HYPO will em­ phasize in describing a trumping counterexample, arc the extra similarities, that it shares with the problem that the cited case docs not, specifically, the set differ­ ence of the set of relevant similarities between p and ccx and that between p and cc, (i.e., Sp,ce:r. \ Sp,cc where Sr,ce:r: :::> Sr,cc·)

(2) Partial: a contrary case that has the sante set or some noncmpty subset of similarities to the prob­ lem as the cited case. A partial counterexample shows, some what more weakly tha11 a trumping courtterexam­ ple, that the shared similarities do not always lead to the same outcome as the cited case. The sci; of all cases, PC X ·,cc, that a side .~ can cite as partial coun tcrexarn­ ples to a cited case cc is the set of all cases, c;, such that the outcome of c; favors sand c; shares some of the same factors with the problem as cc.

The salient features of a partial counterexample are the similarities that botl1 counterexample and cited case have to the problem, namely, the intersection of the set of relevant similarities between p and ccx and that between r and cc, (i.e., Sr,cez n Sr,cc where Sr,cez ~ Sp ,cc . )

(3) Boundary: a contrary case which is an extreme example of some factor, fi, that both is a relevant sim­ ilarity between the problem and the cited case and fa­ vored the winner of the cited case. The set of all cases, RC X ·,cc> that a side .~ can cite as boundary counterex­ amples to a cited case cc is the set of all on point. cases, Cj, such that the outcome of Cj favors.~ and there is some factor, fi, such that Cj and cc share fi with the problem, and the magnitude of J; in Cj is greater in favor of .~'s opponent than in cc.

Por a given boundary counterexample, ccx, the salient features arc the factor, .fi, of which the counterexample is a more extreme example than the cited case and the relative magnitudes of fi in the counterexample and the cited case. Specifically, the salient features arc: (Assume that defendant won the cited case.)

By emphasizing that the plaintiff won despite that fact that the counterexample was so much worse for the plaintiff along factor fi than the cited case, a bound­ ary counterexample tends to show that fi is not a very significant factor. (Like human arguers, HYPO uses counterexamples and distinguishing to deal symbol­ ically with the problem of weighting competing factors. See [Ashley 88].)

(4) Potentially trumping: a contrary case that would be a good counterexample with which to trump tl1e cited case if the problem had certain additional fac­ tors. The set of all cases, PTC X, cc, that a side .~ can cite as partially trumping counte~examples to a cited case cc is the set of all cases, Cj, such that the outcome of Cj favors s and there is some hypothetical variant of the problem, pvar, such that Cj is more on point with respect to pvar than cc. ln HYPO, we have identified a number of heuristics for hypothetically modifying a problem situation to generate pvar. See [Rissland 86].

The salient features of a potentially trumping coun­ terexample are the extra similarities between it and the hypothetical variant of the problem (pvar ), namely the set difference of the set of relevant similarities be­ tween pvar and ccx and that between p and cc, (i.e., Spva.r,cea: \ Sp,cc where Spva.r,cez :::> Sr,cc·) These salient features are the difference between the problem situation and a hypothetically modified problem in whidt a side's arsenal of cases to cite is strengthened by the addition of a new trumping counterexample.

As Figure 3 illustrates, HYPO describes a case differ­ ently depending on how it is used as a counterexample. No. 3, shows how HYPO describes the Data General case as a trumping and boundary counterexample (ccx). Here, the problem situation (p) is a variant of the Crown case in which all 7 disclosures to outsiders were subject to confidentiality restrictions (i.e., Fc,own-var = {f 1 f 2 } and M (f l·rown-var) = 7 ) and the case cited for the de­ fendant (cc) is the Midland RoH case of Figure 2.

In t.he first. sentence of No. 3, HYPO cites Data Gen­ eral as a counterexample to trump Midland RoH and emphasizes the extra similarities that led to the opposite result in Data General: namely f 2 · This comes from the above definition of a trumping counterexample's salient features where {!I J2} \ {fl} = {12}.

In the second sentence of No. 3, HYPO cites Data Gencml as a boundary counterexample, emphasizing the fact that. the plaintiff won in Data General even though it was a much worse case than Crown in terms of dis­ closures to outsiders (6000 outsiders versus 7.) When used as a boundary counterexample, the sa)jent feature of Data General is f 1 , because, making the substitutions m formula (3):

Target Case: The salient features of a case that is a target ( t) for a hypothetical variation of another case to be modified (cmod) are those factors in

l'~ \ T·~m.od that arc added, or those factors in Scmod,t whose magnitudes arc changed, to make the hypotheti­ cal rnorc like or more extreme than the target. As Fig­ ure J, No. ·1, shows, when HYPO describes the Data General case as a target for a hypothetical modification of the Crown case, it emphasizes the extra factor (h dealing with restrictions on disclosures) which, if added to Crown, would make Data General a counterexample with which to trump Midland Ross ([Rissland 86] de­ scribes more completely HYPO's posing of hypotheti­ cals.)

To summarize, HYPO generates a variety of descrip­ tions for any given case in its knowledge base depend­ ing on the viewpoint and context of the argument it makes. As the descriptions of the Data General case in Figure J illustrate, HYPO emphasizes various aspects of the case, sometimes focusing on similarities, some­ times differences, sometimes factor magnitudes, some­ times factors significant because they do not apply. Each of HYPO's thirteen implemented factors has an associ­ ated "canned" phrase for describing it. HYPO composes the phrases corresponding to the salient factors into the brief case summaries. ht this way, HYPO tailors descrip­ tions of the thirty cases in its CKB to fit the argument contexts, viewpoints and problem situations. The de­ scriptions compare favorably with those found in court opinions citing the same cases. See [Ashley 87a]. Al­ though HYPO's text generation scheme is primitive, the salience information could be used as inputs to a so­ phisticated natural language generator for planning more elaborately tailored case descriptions.

6 Conclusion

In this paper, the problem of designing a system that can tailor descriptions of knowledge base objects to suit alternative viewpoints and contexts has been examined for a limited kind of discourse: case-based arguments. Determining the features of a past case that are sal.ient when the case is cited as an argument justification de­ pends on the particular facts of the problem and of the other cases in the knowledge base, the viewpoint. of the arguer and the contextual role of the case in the argu­ ment.

IIYPO, an adversarial case-based reasoning program, tailors its descriptions of a case to emphasize those fea­ tures, specifically factors, that arc salient depending on viewpoint and context. HYPO describes cases in argu­ ments from two conflicting viewpoints, that of plaintiff or defendant, and in four contextual roles, as cited case, dis­ tinguished case, coun terexarnple or target case. HYPO makes further decisions about salience depending on the kind of counterexample a case is used as: trumping, par­ tial, boundary, or potential trump. for each of the con­ texts, we have defined the salient features and illustrated how the definitions apply to a past case described in each context.