Combining Case-Based and Rule-Based Reasoning: A Heuristic Approach

In this paper we discuss a heuristically control led approach to combining reasoning w i th cases and reasoning w i th rules. Our task is interpretat ion of under-defined terms that occur in legal statutes (like the Internal Revenue Code) where certain terms must be applied to part icular cases even though their meanings are not defined by the statute and the statutory rules are unclear as to scope and meaning. We describe this problem, known as statutory interpretation, provide examples of i t , describe the need for melding case-based and rule-based reasoning, and discuss heuristics used in guiding reasoning on such problems. We conclude w i th a discussion of our on-going work to model this mode of expert reasoning. 1 I n t r oduc t i on "Statutory in terpretat ion" is the process of t ry ing to determine the meaning of a legal rule by analyzing its terms and then apply ing it to a part icular set of facts. The diff iculty presented to adjudicators, advocates and administrators by this exercise is that cr i t ical terms are typical ly not defined completely (or at all) by a statute. Further, a rule taken as a whole may have unspoken qualif ications and exceptions. Thus one must look outside a statute to other sources of knowledge for clues to its meaning and the meaning of its constituent elements. In part icular, one tries to resolve interpretat ion problems by considering past applications of the rules and terms in question: by examining precedent cases, comparing and con t ras t ing these wi th the instant case, and arguing why a previous interpretat ion can (or cannot) be applied to the new case l [Levi, 1949; Llewellyn, 1960; Tw in ing and *This work was supported in part by the Advanced Research Projects Agency of the Department of Defense, monitored by the Office of Naval Research under contract no. N00014-87-K-0238, the Office of Naval Research under a University Research Initiative Grant, contract no. N00014-86K-0764, and a grant from GTE Laboratories, Inc., Waltham, Mass. Note, in the fullest sense, interpretation also requires consideration of whether a term or rule "should" be applied. In Miers, 1982]. The intepretat ion problem demands that one combine reasoning wi th cases and reasoning wi th rules (statutes). Whi le the need to mix case-based reasoning ( " C B R " ) and rule-based reasoning ( "RBR" ) is a prototypical feature of statutory legal reasoning, other domains also require i t . We believe our approach can be applied beyond the realm of law; in part icular, to extend t radi t ional expert system approaches to "soft" domains that lack a strong domain model. For examples of underdefined terms in a legal rule, consider a section of the statute that governs the assessment of Federal income tax, the Internal Revenue Code (sometimes called just the "Code") . In stat ing the requirements for taking a home office deduction, Section 280A(c) ( l ) of the Code employs such terms as "principal place of business", "convenience of the employer" and use on a "regular basis". Nowhere are these elements defined in the statute; yet some scope must be afforded them in order to apply the statute to particular cases. Whi le the meaning of such phrases is sometimes elucidated by official regulations issued thereunder by the Internal Revenue Service, a clear-cut definit ion (which does not itself use undefined terms) is almost never to be found. Often, the reach of the meaning of such phrases is fundamental ly unclear, varies greatly according to the factual context in which they are used, and defeats precise definit ion by rules. For clues to their scope, practi t ioners rely on previously l i t igated tax cases that have construed these terms. 1.1 T h e I n t e r p r e t a t i o n P r o b l e m i n t h e L a w In statutory interpretat ion, ambiguous terms often this discussion, we leave aside these important normative aspects, which involve reasoning about legislative intent, policy and ethics[Fuller, 1958; Hart, 1958]. §280A(c)(l) states that a deduction may be taken for "any item to the extent such item is allocable to a portion of the dwelling unit which is exclusively used on a regular basis — (A) [as] the principal place of business for any trade or business of the taxpayer, (B) as a place of business which is used by patients, clients, or customers in meeting or dealing with the taxpayer in the normal course of his trade or business, or (C) in the case of a separate structure which is not attached to the dwelling unit, in connection with the taxpayer's trade or business. In the case of an employee, the preceding sentence shall apply only if the exclusive use referred to in the preceding sentence is for the convenience of the employer." 524 Cognitive Models arise from "open-textured" concepts, and these concepts are often the focus of case-based attack. By "opentextured" concept we mean a concept that cannot be defined by necessary and sufficient conditions: one whose boundary is not sharp. Such concepts have been much discussed in jurisprudence [Hart, 1961; Dwork in, 1977] and also in philosophical discussions of "natura l k ind" classes [Wit tgenstein, 1958; Putnam, 1975]. Gardner's recent work [Gardner, 1987], for example, discussed how such legal open-textured concepts give rise to what are known as "ha rd " cases, that is, cases over whose resolut ion experts (judges, scholars, etc.) disagree. Many concepts in domains like the law are open-textured and sometimes even famil iar terms reveal a surprising opentextured l in ing, such as "contract" or " income". Concepts like "due care", which are used deliberately to in dicate a variable standard of behavior, are clearly of this sort. So are "meeting or dealing" and "exclusive use" from the home office deduction rule. Their interpretat ion is the subject of numerous cases. 3 The need to do statutory interpretat ion is not necessarily the result of poor legal draf t ing. Rather it is a persistent problem that resists a legislature's best good-faith efforts at draft ing t ight statutes. Most generally, the persistence is due to the nature of the law and its relation to society; more part icularly, to factual circumstances unanticipated at the t ime of draft ing and a changing legal context [Levi, 1949; Sunstein, 1988]. This was one of the points of one classic discussion of the problem of statutory interpretat ion known as the "Hart-Ful ler debate", between H.L.A. Hart and Lon Fuller in a Harvard Law Review dialogue [Hart , 1958; Fuller, 1958]. There they discussed, among other things, such deep jur isprudential issues as the nature and status of rules and the role of "ought" (normative considerations) in statutory interpretat ion. 4 For instance, a case involving Max Frankel, The New York Times Managing Editor, Max and Tobia Frankel v. Commissioner, 82 USTC 318 (Filed February 28, 1984), addressed the former. Mr. Frankel maintained an office at his home in the Bronx, which he used for reading the morning papers, writing memoranda, clipping materials, and speaking by telephone to his employees, prominent politicians and community leaders. The Tax Court denied that Mr. Frankel met any of the three disjunctive requirements of the statute, (A) , (B), or (C). In particular, the use of the telephone to conduct business was held not to satisfy the meeting or dealing predicate, which was construed to require the physical presence of business contacts. Two famous hypothetical statutory rules from this debate nicely illustrate the problems: (1) "No vehicles are allowed in the public park." and (2) " I t shall be a misdemeanor . . . to sleep in any railway station." Hart and Fuller were concerned with applying such rules to "hard" cases, where the puzzle is to interpret open-textured concepts, like "sleeping" or "vehicle", in light of a statute's purposes. For instance, does a tank which is part of a war veterans memorial statue count as a vehicle? What about a motorized baby carriage or wheelchair? What about a fire engine requiring access to a fire via the park? As for "sleeping", what should we decide about a bum who has obviously bedded down for the night but still has his eyes open? Should the result be any different as to a well-dressed commuter who has clearly dozed off? 1.2 T h e I n t e r p r e t a t i o n P r o b l e m O u t s i d e t h e L a w Al though law is the focus of this discussion of mixed C B R / R B R paradigm reasoning, lawyers are by no means the only ones to combine these two different modes of reasoning. Mathematicians regularly combine reasoning deductively and reasoning w i th examples. Al though sometimes overshadowed by formal definitions and theorems and their proofs, examples, that is, cases, constitute a powerful aspect of expertise [Rissland, 1978]. Polya [Polya, 1965] speaks of the importance of interleaving these two modes of reasoning in the "alternating process" in which one switches to CBR (to find a counterexample) when deductive reasoning stalls and vice versa. The "dialectical" process discussed by Lakatos [Lakatos, 1976] depends crit ically on use of exemplar cases as much as it does on proof analysis. And in AI discovery systems, such as Lenat's A M , examples are a powerful source of control and focus of attention [Lenat, 1977]. Even in medicine, where heuristic rules have formed the core of the current generation of medical expert systems (e.g., rules of diagnosis as in M Y C I N ) , there is a rich body of specific cases of a phenomenon (e.g., particular cases in a particular practice such as "Mrs. Jones, the woman whose problem turned out to be borderline hypertension") which an expert might use, especially in cases requiring judgment calls. Equally important , even though one tradit ionally treats concepts like "hypertension" as well-defined in expert systems, such terms really are not so clear-cut as all that — for a large part of their meaning lies in how they were used in past cases. 2 The Legalistic Ch i ld : An Example of In terpre ta t ion As an example of the need for statutory interpretat ion, consider the case of the "legalistic ch i ld" from the book by the Bri t ish legal scholars Twining and Miers [Twining and Miers, 1982] "Johnny, aged 7, is an only child. In recent months his mother has been mildly worried because he has developed a craving for sweet things and this has affected his appetite at meal times...Then one afternoon she finds that Johnny has gone into the larder and helped himself to half a pot of strawberry jam...she does not punish Johnny but instead says, 'That's naughty. In the future you are never to enter the larder without my permission/ 'What does enter mean, Mummy?' asks Johnny. To go into', says his mother. 'O.K.' says Johnny, relieved that he has got off so lightly. Several incidents then follow. First, Johnny gets a broom and hooks the pot of jam from the larder and helps himself. 'I didn't enter the larder', he says. Next the cat enters the larder and attacks the salmon which mother has bought for a special occasion. Mother, upstairs, hears Johnny hooting with laughter. She comes down to see him standing outside the larder door watching the cat eating the fish. 'I may not go into the larder,' he says." Clearly one of the conflicts between Johnny and his mother concerns the meaning of enter, another is the Rissland and Skalak 525 scope of the rule itself (e.g., is there an unspoken except ion that allows entry in dire circumstances much like that enabling fire trucks to run red lights?). We shall retu rn to this example to demonstrate our computat ional approach to interpretat ion. 3 Synopsis of C B R and M i x e d Parad igm Approaches Case-based reasoning ( " C B R " ) has grown rapidly in the last few years [Kolodner, 1988; Rissland and K ing , 1988]. W i t h i n CBR there are two major classes of CBR that can be identif ied: problem-solving CBR [Hammond, 1986; Kolodner, 1987; Sycara, 1988] and precedent-based CBR [Ashley, 1988; Ashley and Rissland, 1988]. Precedentbased C B R is distinguished by its focus on the use of past cases ("precedents") to justify a solution and explain its rationale 5 . Anglo-American common law w i th its doctr ine of the binding nature of precedent is a paradigm of precedent-based CBR. On the other hand, in problemsolving CBR, the typical focus is on using past cases to find a a detailed problem solution (e.g., a plan, a course of act ion), where the new solution is generated by adapting a previous solut ion. Industr ia l design and planning are paradigmatic examples of problem-solving C B R [Barlet ta and Mark , 1988]. Both types of C B R follow similar steps. Once a new case has been accepted for analysis, CBR proceeds by (1 ) analysing it (e.g., by comput ing features, relations and indices) to retrieve a set of relevant cases from case memory; ( 2 ) f rom these selecting a subset of best cases from which to craft a solution or interpretat ion for the problem case; (3 ) derivation of a solution or intepretat ion complete w i th support ing arguments in the case of precedent-based C B R and w i th implementat ion details in the case of problem-solving CBR; (4 ) testing of the the interpretat ion (e.g., w i th hypo the t i ca l ) or solution (e.g., w i th simulations) w i th an eye to assessing its correctness, strengths, weaknesses, generality, etc.; and (5 ) storing the newly solved or interpreted case into case memory and appropriately adjust ing indices and other CBR mechanisms such as simi lar i ty metrics. Note, in assessing relevancy in Step 1, and all the other steps of C B R as wel l , one must view cases f rom the point of view of the case and task at hand. So, for instance, just because a known case was a landmark case does not necessarily make it impor tant for the present case since the two might not share any relevant similarit ies. Furthermore, in s tatutory interpretat ion the CBR must address the requirements of the statute. It is not enough simply to argue about the meaning of legal concepts; one must tie the arguments to the statute. This latter remark shows why our past work on H Y P O is insufficient As in previous precedent-based systems of our group, HYPO and TAX-HYPO, the key idea is to reason from cases similar to the current case in order to argue for a particular decision in the current case and to justify the reasoning in terms of the past cases. A large part of the effort is on selecting and arguing about the relevancy of cases: showing similarity with supporting cases and distinguishing contrary cases. in itself for modell ing statutory interpretat ion [Rissland and Skalak, 1989]. At this point , researchers have only recently begun to wri te about the integrat ion of C B R w i th other reasoning paradigms [Goel and Chandrasekaran, 1988; Koton, 1988a; Ko ton , 1988b; Marques et a/., 1988; Walker et a/., 1988]. We feel that such mixed-paradigm approaches are natural and shed l ight on both the cognitive skills in volved in such reasoning and on questions of architecture and control of their computat ional models. 4 Heurist ics for M i x e d Parad igm Reasoning In our study of statutory interpretat ion, we have gathered a collection of 30 or so heuristics that we believe experts use for control l ing and interleaving reasoning w i th rules and reasoning w i th cases. These heuristics can be divided into a number of categories: 6 1. Ways to Begin Reasoning 2. Rule-based Near Miss 3. Rule-based Near Hit 4. Ways to Broaden a Rule 5. Ways to Discredit a Rule 6. Ways to Confirm a Hit 7. Ways to Confirm a Miss 8. Ways to Confirm Reasoning: ''Sanity Checks" 9. Ways to Deal with Results Opposite from that Desired 10. Ways to Deal with Failure of Reasoning to Yield a Definite Conclusion 11. Ways to Focus the Reasoners 12. Open-Textured Elements Some of our heuristics, like those in groups 8 and 12, are very similar to those employed by Gardner [Gardner, 1987]. We are currently exploring the use of these heuristics in control l ing a mixed paradigm system, combining case-based and rule-based reasoning, called C A B A R E T (CAse-BAsed REasoning Tool) . The main features of C A B A R E T ' S architecture are (Figure 1): • There are two co-reasoners (CBR and RBR ). t Each co-reasoner is capable of running in a standalone manner. • Each co-reasoner has a dedicated repor te r process that reports the end results of a reasoner and certain aspects of its intermediate processing. • A con t ro l l i ng process uses reporters' observations and its library of control heuristics to decide how the system as a whole and the individual reasoning processes are to proceed. • The ultimate goal for which the system is working (an argument for a given side or a neutral explanation) is specified by the user, as part of the intit ial input to CABARET. A hit refers to the establishment of the antecedent of a rule, on the rule-based side, or the presence of all the prerequisites of a dimension (index), on the case-based side. A miss is the opposite of a "h i t " . Near miss and near hit are discussed below. S26 Cognitive Models C A B A R E T uses a n agenda-based C o n t r o l l e r i n wh i ch the heur is t i c c o n t r o l ru les d i rec t a n d in te r leave the t w o modes o f reason ing by p o s t i n g a n d p r i o r i t i z i n g tasks for each t o d o . T h e c o n t r o l ru les are w r i t t e n i n C A B A R E T ' s Control Description Language, ( " C D L " ) , w h i c h prov ides a v o c a b u l a r y w i t h w h i c h to express ( i ) h igher level control descriptors t h a t descr ibe at a f a i r l y h igh level the s ta te or resu l t o f the case-based reasoner and the ru lebased reasoner, a n d ( i i ) tasks for the C o n t r o l l e r to suggest for each reasoner. For i ns tance , since w h a t an exp e r t , o r C A B A R E T , does i n a p a r t i c u l a r s i t u a t i o n m a y depend on w h a t side is be ing argued for , the C D L has a desc r ip to r for point-of-view. D e p e n d i n g on whe the r the user w a n t s the consequent of a ru le to be es tab l ished, point-of-view m a y be pro ( f o r ) or con the ru le . 7 C h a n g i n g the p o i n t o f v i ew enables e x p l o r a t i o n o f a s i t u a t i o n f r o m var ious a r g u m e n t a t i v e van tage po in t s . Four groups ( # 4-7 above) o f C A B A R E T ' S heur is t i cs concern such arT In tax law, for instance, the Commissioner of Internal Revenue may argue against a statute that gives a taxpayer a deduct ion f rom his gross income, w i t h the taxpayer arguing for i t . g u m e n t a t i v e stances: c o n f i r m i n g t h a t a ru le shou ld f i re , c o n f i r m i n g t h a t a ru le shou ld n o t f i re , b roaden ing the scope o f a ru le (enab l i ng i t to be app l i ed even w h e n i t seems no t to a p p l y ) , and d i sc red i t i ng ( l i m i t i n g the scope o f ) a ru le . These are t r i gge red as fo l lows : For ins tance, i f Ru le 1 has f i red, b u t y o u d o n ' t l i ke some consequence o f Rule 1 , you (and C A B A R E T ) m a y look for ways t o d iscred i t t h a t ru le . C A B A R E T k n o w s , for examp le , several ways to " d i s c r e d i t " a ru le : f ind cases where the consequent was deemed no t to have been estab l i shed , even t h o u g h the ru le f i red; n a r r o w the reach o f the open tex tu red words in the ru le , and so f o r t h . T h e C o n t r o l Descr ip t ion Language also conta ins desc r ip to rs such as: near miss, near hit, open texture, point of view, most on point cases and p r i m i t i v e task d i rect ives , such as forward chain, backward chain, filter cases, confirm a hit, confirm a miss, broaden, discredit. T y p i c a l tasks for the C B R process are analogize t w o cases ( p o i n t ou t the d imensions in c o m m o n and l ike values a long such d imens ions) and distinguish cases ( p o i n t o u t d imens ions no t in c o m m o n or d iss imi la r values a long shared d i m e n sions) [Ashley, 1988]. T h e C D L descr ip tors near miss and near hit are app l icab le to b o t h the R B R side a n d the C B R side. Genera l ly , a near miss is had when a resul t (say, one t h a t you w a n t ) is m iss ing one compo nent in order to o b t a i n . A ru le-based near miss occurs when a l l b u t one con junc t of a ru le can be es tab l ished. A case-based near miss happens when a l l b u t one prerequis i te of a d imens ion ( i ndex ) are present in the case knowledge base. Examples o f rules in the "near m iss" g r o u p are: • If you have all but one conjunct of the antecedent of a rule, and you want the rule to fire, broaden the