Graham Greenleaf BA LLB MACS
Senior Lecturer in Law
University of New South Wales, Sydney, Australia
Presented at the Australian Legal Convention, Darling Harbour, Sydney, August 1989
Expert systems research is part of the `second wave' of artificial intelligence research. The first wave of the 1950s and 1960s was characterised by attempts to create computer simulations of the basic building blocks of human intelligence, such as logical reasoning, natural (human) language understanding, and learning from experience.
When the switch was turned on, we would have the new-age equivalent of a new-born baby -- but what a baby! Although it would know nothing, it would be incredibly intelligent and capable of learning just the way that a human being learns. All we would need to do would be to feed it information in a raw form and let it learn. And since it could process information so rapidly, within a year or two it would match the knowledge of any human.This first wave was something of a dumper. Developing programs which could prove theorems of logic or play chess at a high level did not seem to provide techniques which could be readily adapted to other tasks. At the end of the day these programs were very `smart' at logic or chess, but still stupid when it came to anything else.
The second wave, which is still cresting, is characterised by a rejection of these `generalist' notions of intelligence, replaced by a recognition of `the primacy of expertise' or `the primacy of domain knowledge'. It was recognised that the essential ingredient of intelligence is organised knowledge, and that even before something can be learnt, a lot must be known, so that the new knowledge can be fitted into an existing structure. Furthermore, it was recognised that a principal characteristic of expert behaviour was not so much general principles of reasoning applicable to any field of human activity, but rather detailed concrete knowledge of very narrow areas. The principle characteristic of an expert system, the model of the new form of computerised artificial intelligence, is therefore the emphasis on the `knowledge base', the creation of a systematic, detailed body of knowledge specific to the particular `domain' or area of knowledge with which the program is concerned. A more general term for this field is therefore `knowledge-based systems', which does not have the connotation of matching expert performance in the way that `expert systems' does.
Many are now attempting to ride this second wave to commercial success, in law and elsewhere. Whether its eventual destination is also to bury our heads back in the sand is the subject of this paper. We begin with some consideration of the notion of legal expertise and the types of knowledge-based applications that might therefore be feasible and attractive in law. We then move to a brief explanation of the components of an expert system, the main type of knowledge-based application that will concern us. We conclude with an examination of some of the inherent limitations of the use of this technology because of the nature of legal reasoning, and a suggestion as to an appropriate way in which legal expert systems should be conceptualised.
However we wish to characterise the reasoning underlying the system of precedent, case law reasoning clearly is not a matter of pure deduction. The reasoning by analogy, capacity to generalise and other elements of case law reasoning are obviously part of the skills of a lawyer.
The aspects of legal knowledge and skills discussed so far all emphasise the lawyer's ability to understand and manipulate the formal sources of legal norms. This ability, while important, is obviously not the only ability of a lawyer, and not the only one that we might wish to computerise.
Two features of these knowledge-based applications distinguish them from data-based legal applications. First, whereas the primary components of legal databases are the `raw materials' of legal decisions, cases and statutes, the primary components of knowledge-based systems are representations of legal knowledge, information about the legal system (including cases and statutes) which has been interpreted and structured by the legal `expert' who created the system. Second, whereas data-based systems merely retrieve `raw material' potentially relevant to the particular legal problem which the user wishes to solve, knowledge-based systems apply that `raw material' to the problem, producing an outcome such as an advice or a document which deals specifically with the problem.
Formal advisory systems must attempt to embody at least general domain knowledge, formal knowledge, reasoning skills, interpretative skills and `real world' knowledge. However, the organisational skills, research skills and strategic knowledge are likely to have much less bearing on the development of such systems.
Examples of such systems are SAL (System for Asbestos Litigation), which advises on the settlement of asbestosis claims against insulators, and LDS (Legal Decisionmaking System) which advises on likely settlement figures for product liability cases, both developed by Waterman et al at the RAND corporation using the ROSIE shell. These programs use knowledge about damages previously paid in other cases, factors indicating likely liability of the defendant and contributory negligence of the plaintiff, characteristics of both litigants, and other factors such as the defendant's attorney. A percentage weight is attached to each item and is used to develop an overall suggested settlement figure.
Such a `strategic' advising system must attempt to capture not only the knowledge and skills required to develop a formal advisory system, but also aspects of the non-formal, `strategic' knowledge of a lawyer, and a good deal more `real world knowledge'.
Such programs rely principally on capturing a lawyer's organisational skills, and perhaps some strategic skills as well, and interpretative skills are also important in document generators.
Proposals for interventionist retrieval systems focus on the actions of the user of a retrieval system, monitoring search requests and results and intervening to assist the user to construct more useful searches (eg automatic truncation, automatic thesauri), or to avoid common search mistakes, and to present results in a more meaningful way (eg ranking documents in order of likely relevance). The problem domain of such approaches is that of legal language and legal search strategies, not any particular area of law.
`Conceptual' retrieval systems, on the other hand, attempt to model relationships between legal terms in a particular area of law, so that requests by user using certain terms are interpreted semantically so as to retrieve other conceptually relevant documents. The techniques known as `hypertext', or non-linear text presentation, are also relevant. By making legal texts in any particular area densely cross-related (between, for example, statutory provisions, cases, commentaries, and reference works), with the user being able to move between these different sources effortlessly, legal texts become much more efficient, revealing and rewarding to read. Conceptual retrieval systems must capture a detailed knowledge of the conceptual structure (formal knowledge and interpretative skills) of an area of law, but even hypertext systems embody knowledge of the formal sources and interpretation (eg knowledge of legal terms of art used which must be cross-referenced to definitions).
The remainder of this paper will ignore all of these possible types of knowledge-based applications to law other than the first -- `formal' legal advisory systems, which is what most people mean by `legal expert systems', even though this term does encompass strategic advisory systems as well.
The inferencing mechanism may operate according to logical rules such as `if A is true and is B true and (A and B implies C) is true, then conclude C is true'. It may also implement procedural criteria such as the program always requiring a first goal of the system to be specified by the application writer, with the system always seeks to prove that that goal is true first. The inferencing mechanism may also embody certain tactical approaches, such as always testing simple rules first.
A common type of inferencing mechanism, backward-chaining (or `goal directed') production rules, requires a knowledge-base in the form of statements of knowledge expressed as `IF premise THEN conclusion' rules. It operates by requiring an initial goal (conclusion) to be specified. It seeks to prove that that goal is true, first by looking in its knowledge-base for a rule which has that goal as its conclusion. When it finds such a rule it seeks to establish that the premises of that rule are true. If it cannot establish that a premise is true, it first looks for a second rule which has that premise as its conclusion, and it then seeks to prove that the premises of that second rule are true, and so on. This is the `backward-chaining' aspect. When it finds a premise which does not appear as the conclusion to another rule, it has no option but to ask the user for the value (true, false or otherwise) of that premise. The value of this premise, once established, can then be used to prove that the conclusions of any rule in which it appears as a premise are true. Rules are said to `fire' when their conclusions are shown to be true.
Writing legal knowledge in the appropriate formalism is what a person writing an application for a particular area of law must do.
A legal expert system `shell' is an expert system with an empty knowledge base. It consists of the inference mechanism, the corresponding requirements of the knowledge formalism, any developer interface which may be provided, and the user interface. A shell doesn't deal with any particular area of law. The application developer uses the shell to develop applications in particular areas of law for use by end-users.
"there are those who are sure that lawyers, for instance, specifically solicitors, could be replaced by computers and that an expert system, albeit a very big one, could actually do the job better. It is true that finding chains of consequences in laws, and finding where laws contradict each other, are ideal tasks for computers and are often done poorly by humans at the moment."Sergot et al  express a similar attitude:
"The knowledge elicitation problem is almost entirely absent in the formalization of legislation. By its very nature, the law is well documented; its provisions are written down, and where they are not, decisions in previous cases are recorded for future reference."However, some authorities on expert systems such as such as Waterman have ahown a more sophisticated appreciation of the special problems presented by law as a domain for expert systems:
"... this characteristic of the [legal] domain, having rules that already exist, has led to trouble. The problem this creates is the naive notion (for some) that because a body of rules and regulations exists, all one has to do is to translate them into executable code to create a program for performing complex legal reasoning."Waterman et al, in the jargon of expert systems research, suggest certain characterisitics of law that make it a more than trivial task to develop legal expert systems:
are rooted, ultimately, in the empirically-based, causal, descriptive laws of the natural sciences, whereas legal reasoning involves the manipulation of the prescriptive laws of the legal order, discoverable, in the main, not from uniformities or patterns in the external world but through scrutiny of the formal sources of the law.The descriptive laws of the natural sciences are a model of the world of physical objects and processes. The formal representation of that model, whether written down in books, or in the mental heuristics of experts, can be considered as independent of the subject-matter that is being modeled. Nor is a statement of a physical law seen as the source or cause of that law. Prescriptive laws, on the other hand, are not a formal model of some independent subject matter: they are the source of legal norms. These legal norms are the subject matter of legal expert systems which simulate formal legal reasoning. Primary legal texts, legislative documents and reports of cases, are, in themselves `just texts', but when interpreted and applied they produce legal norms. As many have stressed, it is the process of interpretation of text which is crucial to the production of norms. Legal reasoning is irreducibly based on language and its interpretation, whereas most other expert system domains are largely concerned with causal relationships between physical objects and processes.
Irrespective of the inferencing mechanism(s) employed, a user's experience in using a legal expert system to solve a problem will not be an experience in pure deduction.
The facts of specific problems do not apply themselves to the legal categories appearing in the language of legal texts, nor in their reflections in the language of questions in legal expert system dialogues. Applying the `rules' set out in legislation to specific fact situations is difficult even where the facts are not in dispute. Pearce  estimates that 40% of recent reported cases required the courts to rule on the meaning of legislation. The well-known causes of ambiguity in legal language guarantee this: legislation usually deals with a wide range of human conduct; it is not possible to foresee every variation of human conduct which might arguably come within a provision; there is no single `natural' meaning of words; legislation is presented in a formal style devoid of examples which may clarify meaning, and so on. This endemic ambiguity must be reflected in the questions which an expert system asks the user in order to elicit the facts of a problem. Bing calls this the `reality interface'.
Effective use of most legal expert systems therefore requires at least some expertise in legal interpretation. The Courts have, of course, developed `rules' of interpretation concerning how Courts may determine the ordinary meaning of words, and when they may depart from those meanings. It might seem that this only raises a slight additional complication: to treat these `rules' and principles of interpretation as a set of meta-rules which need to be formalised and included as a separate component in the expert system. In fact, this is only likely to be achieved to a limited degree. Pearce warns:
It is important at the outset to stress that these are nothing more than approaches and presumptions. To elevate them, as is so often done, to the level of `rules' is but to mislead as it invites the assumption they will be strictly applied by the courts ... the so-called rules can only be regarded as aids to interpretation. Cross describes the principles of interpretation as "a congeries of principles capable of pointing in different directions and incapable of arrangement in any kind of systematic hierarchy according to their different degrees of persuasiveness." If they are correct, then to look for generally applicable set of meta-rules for statutory interpretation is a forlorn search. The skills of legal interpretation cannot, it seems, be reduced to a set of meta-rules.
Susskind provides a thorough discussion of these matters in his rejection of `the argument from particularity of facts' and `the argument from open texture and vagueness' and correctly rejects them as arguments against the development of expert systems based on deductive reasoning, but does not draw the conclusion that developing techniques to assist users to make the interpretative decisions that they are inevitably required to make should be a central feature of legal expert systems research. He reaches the pessimistic conclusion that research should concentrate `on designing systems to solve clear and deductive cases'. Instead, it may be that the key practical question in developing legal expert systems is in finding the dividing line between what an expert system can do (given existing technology) and what elements of solution of a problem the user must provide during a dialogue with the system.
Each does what (s)he or it does best, then hands back control to the other. The program controls those steps in the solution process that involve deductive steps such as deduction of consequences of user-supplied facts and insistence that all aspects of a problem, no matter how minor, be considered, by presenting appropriate questions to the user. It also applies such non-deductive reasoning techniques as may be developed from time to time, and are shown to be reliable in relation to particular types of legal problems. The human user controls those steps of the solution process which involve abilities which cannot (at least as yet) be reduced to a computerised form, including all of the various interpretative skills which lawyers must exercise.
The second vital task of the program is to provide aids to the user's exercise of his or her interpretative skills. Such aids to interpretation include: warnings when terms used in questions or explanations are `terms of art' defined elsewhere in a statute (or in cases); the provision of such definitional material on request, the provision of full text searches of statutes so as to find other contexts in which the same words have been used; and the use of the facts of a problem to instantiate statutory language, and to remove all redundant information from representations of sections, so as to reduce potential misunderstanding cause by the abstract and multi-purpose nature of statutory language.
If this `interpretative' model of a legal expert system in use is adopted, it can indicate many of the features needed in a shell suitable for the development of legal applications.
Such an approach seems to have informed the work of the Knowledge Based Systems Group in Sweden, who have developed PLUTO, a system to assist social workers handling social assistance applications.
`A tool like PLUTO tries to encompass the whole work process. It is not primarily the skills of the expert that are represented in the tool but knowledge of how information is processed in the domain... If needed, the system processes information in order to make the user able to apply the essence of his or her professional skill on the problem at hand.'Ronald Stamper takes a similar approach:
Our goal should be to make legal knowledge less obscure, more clearly structured, more open to the user and the critic. We have less need for a cognitive legal machine than for a less sophisticated but more humble product to support intelligent human interaction.'
Seen from this perspective, the task of developing legal expert
systems is feasible, useful and still just as challenging.
 In addition to journal articles too numerous to mention, reference may be made to R Susskind Expert Systems in Law Clarenden Press Oxford 1987, A vdL Gardiner An Artificial Intelligence Approach to Legal Reasoning MIT Press 1988, P Capper and R Susskind Latent Damage Law -- The Expert System Butterworths London 1988 , C Walter (ed) Computer Power and Legal Reasoning West Publishing Co 1985 and C Walter (Ed) Computer Power and Legal Language Quorum Books NY 1988
 P Capper and R Susskind Latent Damage Law -- The Expert System Butterworths (UK) 1988 contains the system on two disks. It is intended to demonstrate `the potential of expert systems in law', and the authors do not intend that it should be relied upon.
 H Gardiner The Mind's New Science Basic Books 1985, Chapter 6 surveys the relationship between artificial intelligence and expert systems.
 Newell and Simon's Logic Theorist (LT) of 1956, discussed in Gardiner loc cit.
 A recognition attributed to Feigenbaum in 1967, first implemented in his program DENDRAL (1971), which identified chemical compounds from mass spectrograph data at expert levels of performance: Gardiner loc cit.
 see D A Waterman et al `Expert systems for legal decisionmaking' Proc. Second Aust. conf. on Applications of Expert Sytems, Institute of Technology, Sydney, 1986
 see Michele Asprey `the computer as a Robot Lawyer' in the proceedings of this Convention
 Some of this literature is discussed in G Greenleaf, A Mowbray and D Lewis Australasian Computerised Legal Information Handbook Butterworths 1988, Chapter 3.
 D Michie and R Johnson The Creative Computer Pelican (UK) 1984
 M J Sergot
 Waterman et al op cit
 ibid, paraphrasiing pgs25-7
 R Susskind `Expert systems in law: a jurisprudential approach to artificial intelligence and legal reasoning' (1986) 49 Modern Law Review 168
 see, for example, J Bing `Uncertainty, decisions and information systems' in C Ciampi (Ed) Artificial Intelligence and Legal Information Systems, Nort-Holland, 1982
 Most notably by Susskind and P Leith, as to which see Susskind Expert Systems in Law, op cit pgs237-9 and references cited therein.
 see Susskind op cit pgs181-193
 D C Pearce Statutory Interpretation in Australia (2nd Ed) Butterworths, Sydney, 1981, p1
 R Cross Statutory Interpretation Butterworths, London, 1976; Pearce op cit; J Stone Precedent and Law, Butterworths, Sydney, 1985 esp. Ch. 3
 J Bing `The Concept of Rule in Law', in Buchberger, Göranzon and Nygaard (Eds) Artificial Intelligence - Perspectives and Implications, Complex 11/87, Norwegian Research Centre for computers and Law / Universitetsforlaget, Oslo, 1987, p124
 op cit p5
 op cit p29
 contra Susskind, who seems to take an optimistic view of this task, op cit p107.
 op cit p192
 See G Greenleaf, A Mowbray & A Tyree `Communications Aspects of Legal Expert Systems -- incorporating them in shells for lawyers' Proc. National Conference on Law, Computers & Artificial Intelligence University of Exeter, UK, 1988 for a discussion of the use of these techniques in the LES and X-SH shells.
 See G Greenleaf, A Mowbray & A Tyree ibid for an initial attempt to catalogue these features, and to explain them in terms of the LES and X-SH shells.
 A Berg, L Hard and P Docherty, `Development of a KBS support system for handling social assistance', mimeo, The Swedish Agency for Administrative Development (SAFAD), Stockholm
 From a review of A vdL Gardner's An Artificial Intelligence Approach to Legal Reasoning op cit