® InfoJur.ccj.ufsc.br
Daniel Poulin[1], Paul Bratley[2],
Jacques Frémont[1] and Ejan Mackaay[1]
[1]Centre de recherche en droit public
poulind, fremont, mackaay@droit.umontreal.ca
[2]Département d'informatique et de
recherche opérationnelle
bratley@iro.umontreal.ca
Paper submitted to the 4th International Conference on Artificial Intelligence and
Law
to be held in Amsterdam, 15-18 June 1993
Abstract
Most rule-based expert systems have been constructed on the premise that a legal provision ought to have only a single interpretation in the knowledge base. Yet lawyers know that legal concepts and rules often can be given more than one meaning. This paper proposes a model to accommodate such plurality of meanings in an expert system. It draws inspiration from writings on statutory interpretation. The idea of a plurality of meanings conflicts with the principle of isomorphism between legal provisions and knowledge base rules. Isomorphism has much to commend it, in particular for the maintenance of the expert system when the law changes. To accommodate several, possibly conflicting interpretations of the same provision while maintaining the advantages of isomorphism, our model relies on metaknowledge. Besides the usual or lower level rules, which translate substantive law, the system has metalevel rules delimiting the set of these lower level rules that can be called upon during the inference process. This allows the system to produce different and possibly contradictory interpretations of a statutory provision, each of them complete and internally coherent.
Daniel Poulin, Paul Bratley, Jacques Frémont and Ejan Mackaay
A legal expert system expresses as a set of formal rules the norms found in the provisions of a statute or regulation, in case law or other legal texts. The process of constructing the system involves interpreting these legal texts and recasting each of them into one or more formal legal rules. The primary sources of law or law-formulations, to use Susskind's term [Susskind 87, 36-37, 124], must be transformed into law-statements--statements about what the content of the law is--and these in turn must be translated into a formal language as rules of inference which Susskind terms legal productions. These transformations take place outside the context of specific cases, that is without reference to concrete legal problems those rules are designed to solve.
The construction of a legal expert system involves two conceptual steps: to identify the legal norms that a statute conveys and to express these norms as formal rules. Several strategies have been proposed for this transformation process. Some consist of a rather straightforward transposition of law texts into formal language at the expense of a substantial loss of meaning of the legal concepts involved; others call for a subtler and more complex legal analysis. In all cases, however, what is at stake is the interpretation of legal documents.
The problem of interpreting of legal documents is well known to lawyers. In legal usage, the term interpretation is employed in situations where the meaning of a legal text, typically a statutory provision, has to be assessed in a concrete situation. In constructing expert systems, one must interpret legal documents ahead of such concrete applications. This interpretation can only be provisional and the best one can do in constructing the knowledge base is to foreclose as few as possible of the meanings for particular provisions one may ultimately want to consider in concrete situations. Conversely, when using the system in a particular situation, one may want to consider different interpretations and their consequences before committing oneself to any one of them. A knowledge base should therefore be designed to accommodate several different meanings of any one provision and not be limited to the one that looks most plausible at the time of construction. One might term this a one-to-many representation of knowledge.
A one-to-many representation of knowledge is quite hospitable to the open texture of the law. The precision of most statutory texts may convey to the non-lawyer the idea that legal concepts and provisions are unequivocal and that the only real difficulty is to design a procedure to determine the meaning. For expert systems design, this leads quite naturally to a one-to-one conception, in which each concept has a single meaning and to each statutory provision (law-formulation) corresponds precisely one formal rule (legal production). Lawyers know that such a conception is illusory. Legal concepts are never fully determined. The natural language terms in which the law is expressed are invariably multifaceted. Translation by a single term is likely to be an impoverishment, even though one uses the same term as the legislator. What will be used in the inferences of the expert system are not natural language concepts but symbols belonging to a formal language with precise semantics. These symbols have only a single meaning.
To move to a one-to-many representation of legal knowledge, one would hope to draw inspiration from the literature on legal interpretation. Much of this literature has been written by scholars of legal theory interested either in justifying judicial decisions or in proposing reforms. More useful for expert system designers are works, usually aimed at practitioners, that describe in great detail how one goes about interpreting legal texts and what canons of interpretation are used [Côté 91; Maxwell 69]. Even these works, however, have not been written with a view to designing expert systems.
A good starting point for expert systems design is the research by the "Comparative Statutory Interpretation Group (CSIG)." This research has taken place over the years 1983 to 1990 under the leadership of MacCormick and Summers. The purpose of this research was to study interpretative practices in nine countries. The results have been published [MacCormick 91]. This publication is our starting point.
A major conclusion of this study is that interpretative practice in the nine countries surveyed, beyond surface differences which can be rationally accounted for, shows substantial similarities. These similarities are evident in the type of argument used. Eleven kinds of argument are identified which can be classified in four broad categories [MacCormick 91, 512 ff.]: linguistic arguments (the ordinary and technical meaning of words), systemic arguments (contextual-harmonization, argument from precedent, analogy with other statutory provisions, logical-conceptual argument, general principles of law and arguments from history), teleological/evaluative arguments (the argument from the purpose or from substantive reasons) and what the authors term "transcategorical" arguments (interpretation in conformity with legislative intent, where it can be identified).
The arguments are used in the reasons for judgment. The expression of the reasons varies widely: from a simple subsumptive model through complex or sophisticated subsumption to the discursive alternative justification, "in which the final decision is not presented as a logical consequence of given premises but as the outcome of judicial choices made according to arguments and priority rules" [MacCormick 91, 493]. In countries where this last form of justification is practised the use of different types of argument can be most easily observed. MacCormick and Summers formulate a model identifying general patterns of interaction among the different categories of argument.
This model is based on earlier work by Wróblewski [Wróblewski 84; 88]. Wróblewski's idea is that the justification of a decision rests on three elements: a set of facts and primary legal rules; a choice justifiable in terms of rules of interpretation; a set of higher level rules (interpretative meta-rules) and values, which control the application of the rules of interpretation. The work of MacCormick and Summers deals with the contents of such rules of interpretation and higher level rules.
MacCormick and Summers propose an ordering of methods of interpretation. They summarise their model thus:
"(a) In interpreting a statutory provision, consider the types of argument in the following order:
(i) linguistic arguments;
(ii) systemic arguments;
(iii) teleological-evaluative arguments;
(b) Accept as prima facie justified a clear interpretation at level (i) unless there is some reason to proceed to level (ii); where level (ii) has for sufficient reason been invoked, accept as prima facie justified a clear interpretation at level (ii) unless there is some reason to move to level (iii); in the event of proceeding to level (iii), accept as justified only the interpretation best supported by the whole range of applicable arguments.
c) Take account of arguments from intention and other transcategorical argument (if any) as grounds which may be relevant for departing from the above prima facie ordering." [MacCormick 91, 531-532]
This specification provides a suitable starting point for the formulation of rules and their representation and use by an expert system on statute law. We do not plan to follow the MacCormick/Summers model in all details. Their objective, to account for decision-making and interpretation by the highest courts on the most difficult legal questions, is much broader than ours; the model we develop produces no novel interpretations. Of the MacCormick/Summers model we retain the idea of hierarchy and ordering amongst interpretative rules. Their detailed description of arguments and their modelling of conflict resolution amongst them will be left for future research. In what follows, we set out a model for representing in a coherent fashion interpretations produced by cases, in scholarly writing and by experts.
An expert system incorporating interpretative models requires new architectures for the knowledge base. Whatever solution is adopted must allow the characteristics of statutory law to be represented correctly. Two current computer paradigms must be combined if the system we envisage is to be implemented cleanly. First, certain elements of the so-called "isomorphic" approach are essential for a viable system. However, this approach must be modified by the addition of a supplementary layer of metaknowledge capable of representing the interpretative model to be implemented.
The basic idea of the isomorphic approach is to take advantage of the formalism inherent in legal texts to translate them and formalize them further as executable logic programs [Sergot 86; Bench-Capon 87]. On this view the knowledge engineer follows the legal sources as closely as possible, writing logical propositions that simply paraphrase or reformulate items in the source text. Items are translated quite literally: the sense ascribed to an item is the straightforward reading that would seem evident to ordinary persons, and this straightforward sense is converted into a logical proposition.
This approach is exemplified by the work of Bench-Capon and Coenen. These authors rigorously respect the structure of the source texts. For them, "even an inconveniently structured piece of legislation should have its structure respected" [Bench-Capon 92, 75]. To justify this, they claim at least four advantages for the isomorphic approach: for development, the approach "imposes a discipline relatively easy to follow" [p. 70] and allows complex texts to be partitioned into convenient pieces; for validation, it makes it possible to verify the rules systematically; for maintenance, when the law changes, the isomorphic approach facilitates identification of the changes that must be made in the knowledge base; and in use, it is easier to follow the chain of reasoning covered by the system.
Of these arguments, the most convincing are those concerning validation and maintenance of the system. It seems evident that strict application of the isomorphic approach would allow the construction of systems that are easy to validate and to maintain, precisely because of the close correspondence between items in the legal text and rules in the knowledge base. Unfortunately, however, this claimed advantage has so far remained purely theoretical, since no useful system has been constructed along purely isomorphic lines.
Recently, Kowalski and Sergot have discussed the benefits that could accrue from using a richer representation of the legal texts . They remark:[[dieresis]]"Since our programs can only indicate what follows from one precise interpretation of the law, we need to consider whether an expert lawyer would find such an ability useful." [Kowalski 90, 212]. While systems using a knowledge base that corresponds to a straightforward reading of the legal texts may be of some utility in the routine application of the law within government departments, assisting more sophisticated lawyers requires less limited systems. This leads Kowalski and Sergot to advocate systems capable of incorporating more than one interpretation, and including a mechanism for maintaining consistency. In their opinion, "These mechanisms can be implemented by metalevel reasoning, by reasoning about the alternative logical models and how they are related to one another" [Kowalski 90, 215]. Routen and Bench-Capon [Routen 91] also envisage using rules at a metalevel to represent facets of the legal texts difficult to reconcile with the usual techniques of logic programming: exceptions, "deeming provisions" (as when X is `deemed' to be Y), counterfactuals, and so on.
The idea underlying multilevel architectures is to distinguish between basic knowledge that concerns the field to be modelled (the object layer), and knowledge that concerns the items in the object layer (metaknowledge). Thus the system has both knowledge about the field of application, and knowledge about this knowledge, or metaknowledge. This idea is not new. As early as 1983, Lenat et al. pointed out how such metaknowledge can be useful: "If metaknowledge is going to be present (and it always seems to be desirable), the knowledge engineer should deal with it explicitly. [...] As a result, the program will function better and will also be easier to build and to modify" [Lenat 83, 222].
To these advantages, van Harmelen adds that such a system allows the same object level knowledge to serve several ends; that the system is better able to produce explanations of its reasoning; and finally:
One of the main purposes of having a meta-level architecture at all is to allow the object-level to be purely declarative, without having to worry about procedural aspects. Thus, for any given query, the object-level (implicitly) specifies a set of answers. It is the task of the meta-level interpreter to determine which of these possible answers is going to be actually computed, and in which order. [Van Harmelen 89, 115]
Despite the interest of such a design, few systems using metaknowledge have so far been presented in artificial intelligence and law. Gardner [Gardner 87] uses metaknowledge in a straightforward way: it serves to control the inference mechanism, to avoid unnecessary inferences and to introduce hypothetical knowledge. In Prolexs, metaknowledge represented as a classification network provides procedural knowledge [Walker 91].
Closest to our proposal is Hamfelt's system [Hamfelt 92], which uses several levels of metaknowledge. He claims that a straightforward translation of the statute text unduly abridges the scope of the legal rules in the statute. He therefore proposes formalizing the principles that allow the law to be interpreted. In his system, legal rules are represented by rule-schemes, from which metarules produce the rules to be applied in concrete instances. It remains to be seen whether metarules producing acceptable legal interpretations can be spelled out.
Starting from the interpretative models mentioned previously, we propose to design a system that reflects more faithfully than previous designs the nature of legal reasoning, in that it allows the possibility of differing interpretations of the law. We are convinced that much is to be gained if the isomorphic approach is enriched by the use of metaknowledge and by incorporating a rich interpretative model.
In the areas which concern us, the law changes frequently [Bratley 91], and this `volatility' cannot be ignored. So far only the isomorphic approach offers hope of mastering this volatility, but it has the disadvantage of permitting only an oversimplified representation of the law. Nor is it possible using the isomorphic approach to introduce into the rule base other knowledge necessary when the rules are applied, such as heuristic knowledge gleaned from experts, or their procedural expertise. In the system we propose, the use of metaprogramming allows heuristic and procedural rules to be represented separately, leaving us free to express the substantive rules in a fashion that links them to the statute. This separation of the substantive rules coming from the legal text and the other elements of knowledge used in legal reasoning promises to facilitate the maintenance of the system when the law changes.
In the system proposed, a single provision in a legal source may be represented by several rules in the knowledge base. A provision whose meaning is plain and uncontested is represented by just one rule. Provisions that can be interpreted in several ways give rise to as many rules in the object level knowledge base as there are defensible interpretations. Each rule is tagged with a label saying which particular rule of interpretation justifies its inclusion in the knowledge base. It may carry further tags, such as a reference to the corresponding statutory provision, perhaps some kind of priority expressing how confidently this interpretation can be sustained, and so on. There is thus a many-to-one relationship between rules in the knowledge base and statutory provisions.
The question of course arises of the coherence of the rules in the object level knowledge base. What we propose allows several meanings to be represented for a single provision of law. Should we not therefore fear that the system will contain contradictory rules, and collapse? Fortunately such fatal contradictions can be avoided. For while representing the full semantic richness of the source texts means including contradictory rules at the object level, yet when a chain of reasoning is constructed, the metarules controlling the inference process can be designed to avoid appealing in the same argument both to a rule and its opposite. Thus each individual argument produced will be coherent. Nevertheless by invoking different rules at the object level, the metalevel rules may be able to construct complete and coherent arguments in support of a particular conclusion as well as against it.
A lawyer uses many types of knowledge, substantive, interpretative, procedural and common sense. If these different types of knowledge are tumbled pell-mell into a single mass of rules, confusion follows. Using several levels of knowledge allows the situation to be represented much more clearly. While the first level of the knowledge base holds the substantive rules stemming from the legal texts in question, the second level represents the lawyer's other knowledge. Rules at the object level concern actions to be taken and states of affairs in the world outside the system; the metarules, on the other hand, concern actions to be taken and states of affairs regarding the rules in the object level. Metarules impose a certain order on object rules to impose coherent readings of the law.
From the computer scientist's point of view, variables in the metarules can be bound not only to objects in the world outside the system, but also to rules or even sets of rules from the object level. When the metarules call the inference mechanism, they will pass to it a subset of the object rules, chosen using the tags assigned to the object level rules. The subset may, for example, include only the object level rules acceptable in a particular style of interpretation. As mentioned earlier, the subset will be chosen, too, so as to avoid including contradictory object level rules that would allow the inference mechanism to generate unacceptable conclusions. Thus instead of leaving the inference mechanism free to find its own way through the rules in the object level, the metarules confine it to only a coherent subset of object level rules corresponding to a particular legal point of view.
The legal metaknowledge to be incorporated is of four types: general, procedural, adversarial and inferential.
a) General The interpretative models mentioned above can be expressed formally in the metal evel and thus be made available to guide and control the inferences made by a legal expert system. They can be used in a general way to provide a framework for the inference process and to reduce the search space. In particular, they can be used to control the inference process when the rule base includes contradictory formulations of the judicial constraints. They also serve to weigh the persuasive force of arguments: one may be based on the most likely interpretation, another on a less likely but still plausible one, and so on. Finally the interpretative models implemented in the metalevel are invaluable as an aid to producing explanations and justifications of the legal arguments thus constructed.
b) Procedural Applying a legal text requires more than the ability to produce one or more acceptable interpretations. A legal expert system must also include representations of procedural and methodological knowledge related to particular situations. The lawyer knows, for instance, that he must first determine whether the current situation falls into some particular class, whether some particular provision of the law applies in this case, and so on. Knowledge of this type reduces the search space, and helps organise the dialogue with the user. However it cannot conveniently be incorporated within the object level rules. In the isomorphic approach, the principle that the structure of the object level rules should correspond to the structure of the legal texts provides no guidance as to how these rules should be used.
Several advantages arise when procedural knowledge is separated from the representation of the rules. Aiello [88, 246-247] points out that if the way the rule base is to be applied is encoded explicitly at the object level, then this single use is the only one possible. It is preferable to maintain the generality of the rule base. However `general' rules cannot easily be used in practice unless the system incorporates procedural knowledge, represented here at the metalevel. Furthermore a rule base free from control information, designed straightforwardly to represent the judicial provisions and structured in the same way as the legislative text, is easier to modify as the law evolves than one where other considerations, such as control of the inference mechanism, complicate the meaning of the rules.
c) Adversarial Much as for procedural knowledge, we propose using metaknowledge to represent specific strategies of argument that can be employed depending on the point of view of the user. In a legal knowledge base representing only one interpretation of the law, such `points of view' are unthinkable, and whatever line of argument the user wishes to adopt, the result is the same. In the system we propose, the presence of alternative rules and of differing interpretations allows us to produce different inferences starting from the same facts. Flexibility at the metalevel is employed to produce inferences favourable to one or the other of the parties involved.
Consider, for example, the law concerning benefits paid by the government to individuals. It is to be supposed that the people administering the law favour careful use of public funds, so they will try to avoid giving money to anybody not entitled to it. On the other hand, voluntary associations whose aim is to defend the beneficiaries of those payments will try to maximize the amounts paid to their members. Although both parties found their arguments on the same statutory texts, there is frequently a considerable difference between the interpretations of these texts that they advance. Moreover often, if not always, the lawyer for one of the parties, even if he is convinced he should win, does not stop his examination of the case on discovering the first argument he can use in his favour. On the contrary, even when he believes his case to be strong, he will want to know what counterarguments can be advanced against it, so he may prepare a rebuttal. The rules of inference at the metalevel must implement a similar strategy for using the object level knowledge.
In fields where the law is determined by case law rather than by statute, a number of researchers have described systems able to reason for both sides of a question. Such systems usually employ case-based reasoning [Rissland 87; Ashley 90]. Ashley, for example, describes the interest of such systems thus:
Designers of expert systems in all fields need to focus on building systems that present alternatives. To many, a computer is a machine that accepts a question and returns the answer. [...] That model does not work in a domain where there is no right answer and where the most important function of a computer is to assist, not supplant, the decision maker by providing a range of alternative reasonable courses of action. [Ashley 90, 4-5]
However in the area of statute-based law little work has been done on such systems. Rissland and Skalak have described the CABARET system, a combination of a rule-based and a case-based system, able to argue either side of a question [Rissland 89; Skalak 92]. In their opinion, case-based reasoning is preferable for a faithful representation of incompletely defined concepts and of the multiple interpretations to which these can give rise. This, they observe, is because when faced with indeterminacy in the law, only the courts can settle the issue:
In particular, one tries to resolve interpretation problems by considering past applications of the rules and terms in question: by examining precedent cases, comparing and contrasting these with the instant case, and arguing why a previous interpretation can (or cannot) be applied to the new case. [Rissland 89, 46]
d) Inferential Experienced lawyers know how to cope with the ambiguity inherent in legal rules. A legal expert system must therefore also be able to function coherently in the face of ambiguity. First, it must be capable of giving a plain answer based on the most evident, straightforward reading of the law. In the field mentioned above, concerning government benefit payments to individuals, it is often clear to everybody involved that X is entitled to such-and-such a payment, and that Y is not. Even when the knowledge base incorporates different interpretations of the statutory rules, such plain answers must be possible. Hence the system must know which rules represent on the face of things the most evident readings of the legal texts. It must also be able to build alternative chains of reasoning in cases where no single, obvious answer can be found, and to produce an argument supporting a predetermined point of view. To achieve this, the system must be provided with metarules that can identify the points in an argument where a different course might have been taken, and that can indicate the consequences of following the alternative path. With such a mechanism, the system will be able to produce alternative inferences leading to different conclusions. Finally it must be capable of explaining and justifying the inferences it has made in any particular case in terms of the interpretative model used.
The idea of using theories of interpretation in expert systems to capture more of lawyers' reasoning is not new with us. Bing [91] deplores that no effort has been made to represent in legal knowledge bases the internal legal rules operating on the substantive rules. Others, while relying on interpretative theories, have sketched systems that could accommodate some elements of such theories. Gordon [91] proposes a <<Truth Maintenance System>> capable of accommodating "alternative" interpretations of legal texts. Prakken [91] wants to use logic to model disagreement on legal questions. Oskamp [89] proposes to use metaknowledge to co-ordinate the use of rules from different sources: expertise, case law, scholarly writings and legislation. Yoshino [92] describes processes that apply validity and priority constraints on substantive rules.
Several researchers have drawn on theory developed by legal scholars to design systems mimicking operations particular to legal reasoning Gardner [87], McCarty [83; 91] and Branting [91]. This is not to say that expert systems developed along more traditional lines take no account of legal theory. The Nervous Shock Advisor [Smith 87], Latent Damage Advisor [Capper 88] and Mairilog [Bourcier 90] are there to show the contrary. But neither the first group nor the second specifically incorporate an interpretative model in their designs.
Closest to what we propose here is the work that explicitly deals with the interpretation of legal concepts or that relies on metaprogramming to reflect interpretative rules. Rissland and Skalak use cases to produce arguments about the interpretation of statutes [Rissland 89; Skalak 92]. They propose a case analysis method and arrive at indices of similarity between a hypothetical case and those already known to the system. From these similarities it is possible to construct different arguments of what the law is in a given case. Ashley [90] takes a similar approach in his HYPO system. Hamfelt [90; 92] explicitly incorporates interpretative rules and metaknowledge into his system for interpreting legal rules.
Constructing the knowledge base for an expert system involves the interpretation of legal texts. This interpretation, while taking place ahead of concrete legal cases, is not fundamentally different from the kind that lawyers engage in during lawyering. When lawyers raise the issue of the interpretation, they usually wish to argue that several interpretations can plausibly be given to a legal text.
Expert systems designs have not so far been hospitable to this view. Most rule-based systems are constructed on the premise that for any statutory provision or set of provisions there should be only a single formal rule in the knowledge base. Allowing several formal rules to correspond to a single set of legal source material would lead to incoherence. In many systems, the "single translation" view is combined with the isomorphic approach, which holds that the source material should be chosen in units coinciding as much as possible to single statutory provisions. The isomorphic approach is particularly helpful for the maintenance of expert systems when the underlying law changes.
This paper proposes a model to accommodate in an expert system a plurality of meanings for a statutory provision or other unit of legal source material, while maintaining some of the advantages of isomorphism. The model draws its inspiration from work by MacCormick and Summers on interpreting statutes, which recognises hierarchy and order among rules of interpretation. In our model, the expert system uses two kinds of rules: the usual or lower level rules, which translate substantive law, and metalevel rules delimiting the set of these lower level rules that can be called upon during the inference process. This allows the system to produce different and possibly contradictory interpretations of a statutory provision, each of which is in itself complete and coherent.
We envisage several kinds of metalevel rules: general, procedural, adversarial and inferential. They correspond to different kinds of intuitive knowledge lawyers draw on. The introduction of metarules not only allows us to accommodate several, possibly conflicting meaning of legal provisions within a single knowledge base, but also clarifies the explanations the system can provide a user in accounting for its conclusions. The system should nonetheless come to unambiguous conclusions in straightforward cases. While several researchers have discussed metaknowledge or have brought it into their models, none, to our knowledge, has used it to accommodate multiple interpretations of legal norms.