Alan L Tyree

The DataLex Project

Introduction

This paper describes the DataLex Project, a joint research project by the authors which has resulted in the development of a "shell" which is uniquely suited to the development of legal expert systems. The shell has been used to develop a number of experimental applications programs which are currently being used for teaching and demonstration purposes in several Australian universities. We believe that there are a number of characteristics which make the development of legal expert systems different in character from the development of similar systems in engineering, medicine and the natural sciences. A consideration of these differences, together with some practical constraints, has been responsible for the particular form which the software tools have taken.

The project is concerned with building systems which simulate the advice that a lawyer might give concerning a client's legal rights or which simulate the type of argument which a lawyer might put to a court. Some of our applications simulate the decision which a court might give on a dispute before it. In addition, we expect that the DataLex systems will be useful in developing "intelligent" computer aided instruction systems.

These are not, of course, the only functions that expert systems could be designed to simulate. There is an important and very practical role for such systems in the preparation of generating legal documents, in improving the performance of commercial document retrieval systems and in assisting the lawyer with case management and support. Although the software tools developed under the DataLex project may be useful in these roles, they are not the main focus of our research.

Law and Expert Systems

Perhaps it should not be necessary to include a section which describes some of the fundamental problems of building expert systems in law, yet it is clear that there are a number of systems which have been built or are being built by people whose primary interest and training is in computer science and who fundamentally misunderstand the nature of law and legal reasoning.

It is particularly strange that law should suffer in this way, since the standard texts on building expert systems emphasize the importance of capturing the knowledge of an expert in the domain area.1 Yet, people who would not dream of beginning construction of a medical expert system without the firm support of a medical physician will happily go to work on a legal system.

The reason, of course, lies in the widespread misconception that law is a simple system of rules and that legal inference consists of a simple deductive application of these rules. Since most expert systems are rule based, the translation job seems easy. Thus:

There are those who are sure that lawyers.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.2

Or, again, this by a group which has built a large legal expert system:

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.3

Any lawyer could give any number of examples and arguments which would illustrate the extent to which these quotations miss the reality of the nature of law. It is enough to give one example and one argument.

An example: the statute reads "...every person shall..". It is clear that the "correct" representation is "(all x)(if person(x) then ...)". However, the New Zealand Court of Appeal held that the correct representation was "(all x)(if person(x) and not(solicitor(x)) then...)".4 It is important for the non-lawyer to note that this decision did not come as any great surprise to lawyers who were knowledgeable in the field. Any expert could have advised that the naive translation was risky.

An argument: the example would not bother those who believe that "...decisions in previous cases.." are sufficient to guide development of a legislative expert system. However, in the English legal system, the system that the authors above are describing, a decision which interprets the words in one legislative context is not valid authority for the interpretation of the same words in any other piece of legislation.

The truth of the matter is that legal "rules", if they exist at all, result from the interpretation of the formal sources of law, an interpretation which is irreducibly based on language and its interpretation. A great deal of a lawyer's expertise consists of knowledge which assists in this interpretation; a system which fails to include this knowledge is unlikely to be very "expert".

In comparison, most other domains which have been the subject of expert systems are largely concerned with the causal relationships between physical processes and objects. The rules in such a domain are intended to model the physical process, even though the modelling might not necessarily be at the level of fundamental mechanisms. It is the formulation of these rules, possibly with the assistance of a knowledge engineer, which is the exercise of expertise.

This misconception about the nature of the law has not escaped the notice of all expert system builders. In our view, it is not surprising that some of the most sophisticated and successful systems have been built by Waterman, et al.5 They have clearly devoted a considerable amount of time in identifying the nature of the legal domain and have concluded:

...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.6

It is sometimes suggested that the problem of interpretation of statutes may itself be solved by rules of statutory interpretation. It is true that the Courts have developed a number of "rules" of interpretation which assist in defining the meaning of words used in statutes. However, again it is naive to think that these rules may be simply coded in order to solve the problem. They are not really in the nature of meta-rules at all, but rather guidelines which may or may not be used.7 One of the recent Australian textbooks on the subject says:

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.8

We are not claiming that effective rule-based expert systems may not be built in the legal domain, merely that it is no less difficult in this domain than in any other and that a simple minded translation of basic legal sources into executable code will produce a legal expert which is a ripe candidate for disbarment.

Statute Law/Case Law

The matter becomes more complex when it is recognized that much of the law is not written down in a simple formulation but must be extracted from decided cases. In the simplest form, the doctrine of precedent holds that each case contains a rule of law known as the ratio decidendi. This rule of law is as binding on later inferior courts as a rule in a statute book. In this simple theory, it is only necessary to examine the ratio decidendi of those cases which are binding on the court, see if they apply to the facts in the dispute before the court and make the decision accordingly.

Every law student soon discovers the cruel joke which is inherent in this simple formulation. The hard truth of the matter is that the simple theory does not work; judges always have some leeway of choice in the application of case law. It is argued by one of the greatest legal scholars that this is due to the different methods which may be used to find the ratio decidendi of a case and difficulties of determining which facts are material (and therefore directly determine which previous cases might be similar). Indeed, Stone argues that cases simply do not have a single ratio decidendi which will explain the decision and which is discoverable from the decision itself:

For, however much we try to conceal the truth by using singular terms like "case, "precedent", "decision", or "holding", the truth is that the ratio decidendi of a case always has to be sought in a body of judicial discourse, that is, of communications by judges which enter the legal materials as a more or less complex collocation of words in a written report.9

Notice that Stone does not say, as some have done, that past case law is meaningless, only that the interpretation and application cannot be done in a vacuum; the meaning of a case is not to be determined merely from the text of the case but rather from the entire body of legal materials. The meaning of a case, or indeed of a word in a statute, might change over time as the body of basic legal materials change; there is a constant need to interpret and reinterpret fundamental legal source material.

This need for constant reinterpretation and reference to the body of source material is reflected in one very obvious way. The only area of computing in which lawyers have been pioneers is in the use of full text information retrieval systems. It seems very likely to us that any large scale expert system development will necessarily need to have a close relationship to these legal sources, most probably providing the user with direct, even automatic, access to the source texts.10

In summary, the "rules" of law are very much more complex than a simple translation of statutory language into executable code. Any attempt to state the law as a body of rules will necessarily have long and complex predicates which can not easily be defined in more fundamental terms. The rules are likely to be, in some instances, deliberately ambiguous, certainly incomplete and probably contradictory. It is often difficult to say with certainty when a given legal predicate is true of given facts.

Other factors which make legal systems hard to build

Quite apart from the unique subject matter, there are other factors which combine to make legal expert systems more difficult to build than those in other domains.

Because of the above mentioned difficulties of formulating legal rules from the primary sources, lawyers have developed a large body of very informal knowledge which assists in understanding and reasoning with the body of more formal legal knowledge. This knowledge must somehow be incorporated into any interesting legal expert system.

In applying this knowledge to the interpretation process, lawyers routinely use a wide variety of reasoning processes. These range from straight deductive reasoning, which is relatively easy to incorporate into an expert system, through analogical reasoning, particularly when dealing with case law and matters of interpretation, to some form of common sense which relies on a large body of seeming extraneous knowledge.

As an example of the latter, consider, Douglass v Lloyds Bank.11 One Fenwicke had deposited some £6000 with a branch of the defendant in 1868. The account was never operated after 1868. In 1927, some of Fenwicke's survivors found a deposit receipt which showed a balance of £3500. Although they had no other evidence of the account, they claimed that amount from the bank with interest. The Court heard evidence of Fenwicke's careful financial habits and of Lloyds' careful accounting practices. There was also evidence that Fenwicke had purchased a parcel of shares worth approximately £3000 in the early 1870s.

In the course of the judgment, Roche J said:

...I recognize to the full the strength of the fact that the plaintiff produces this deposit receipt, but I cannot ignore what experience tells me, and the evidence in this case shows, that people lose or mislay their deposit receipts at the time when they want to get their money back, and that money is paid over , if they are respectable persons and are willing to give the necessary indemnity...12

It is easy to see that the decision is correct, yet it is not easy to see how to build an expert system which would have come to the same conclusion without the builder actually foreseeing the particular circumstances of the case.

Every expert system builder faces the problem that natural language concepts are ambiguous and that the words may be used in subtly different ways, but because concepts and words are the fundamental building blocks of law, the problem is exacerbated. The concepts used are ordinarily open textured. The same word is often used in different ways, sometimes even in the same statute. For example, the word "negotiable", one of the more technical and well understood legal words, is used in at least two different ways within the Bills of Exchange Act 1882.13

"Deep" vs "Shallow" models

There has been considerable discussion over the need to construct "deep conceptual" models before effective legal expert systems may be built. McCarty, for example, has argued that the characteristics of the legal domain described above mean that some fundamentally different approach to the construction of expert systems is called for.14

The meaning to be given to "deep" and "shallow" is not always clear. There must be at least three levels at which a legal expert system might be expected to incorporate or represent legal knowledge:

the system could include only the heuristics of legal experts as to the outcomes which are likely in particular situations, but without provision of any justification based on primary legal sources;

the representation could include justification based on the primary legal sources, but without any explicit causal model of those sources; certain heuristics concerning the relationship between these sources, eg, principles of interpretation, may be implied in the representation or the inference system;

the system could include an explicit causal model which serves to define the relationships among the concepts employed in the primary sources. Justification would then, presumably, be based on the model.

The "deep" vs "shallow" argument does not seem to us to be a matter for argument from a priori principles. It is certainly possible now to build systems which function at depth (ii) on the above list. We believe that the absence of anything even resembling a "deep" model of more than the smallest subset of the legal domain means that expert systems of type (iii) are far in the future and will require very substantial resources to build. Whether the expenditure of these resources is necessary or even justified, at least for the purpose of building expert systems, will depend upon the performance of level (ii) type systems. That is an empirical problem which may only be resolved by building systems and evaluating their performance.

Knowledge Acquisition

Susskind argues that:

The activity of legal knowledge representation...involves the operation of interpretative processes whereby the legal data of part of a legal system...is scrutinized, analyzed and eventually reformulated in a fashion that is both faithful in meaning to the original source materials and that allows for the requisite transparency and flexibility of expert systems in law.15

With these comments we respectfully agree, adding that there are some very practical consequences which flow from the observation. The traditional method of acquiring knowledge for an expert system makes use of the so-called "knowledge engineer", a person who is wholly familiar with the representations of knowledge which are internal to the expert system and who is skilled at extracting knowledge from the domain expert and translating it into the internal representation.

There are several problems with this process. First, it is very easy for the knowledge engineer to believe that he or she knows a great deal about the domain. This is not an original observation. One of the leading texts on the construction of expert systems advises:16

Be careful about feeling expert.

It is very easy to be deluded into thinking one knows a great deal about the domain. Remember: the expert became one only after years of training and experience.

This problem is magnified in the legal domain, the problem being so common that it even has a name: in Australia such a person is called a "bush lawyer".

The second problem with the "domain expert - knowledge engineer - expert system" paradigm is simply one of time and money. At least in the Commonwealth countries, law schools are not oversupplied with resources. If the knowledge engineer is a necessary component of developing expert systems, then such systems will simply have to be built somewhere else. We will argue below that it is possible to devise representation schemes which will allow lawyers to be their own knowledge engineer.

The DataLex Project

Basic DataLex Design Criteria

1 Direct construction by lawyers: For the reasons mentioned above, we consider it desirable that the construction of legal expert systems is done directly by a lawyer without the use of a knowledge engineer intermediary. This has several important corollaries.

First, most lawyers do not have significant experience with computer languages, nor do they have the time or the desire to learn. The form of representation must be such that expression in that form comes naturally to lawyers. The form should be the simplest possible which is still consistent with the fundamental requirements of legal knowledge representation.

On the other hand, the need for multiple reasoning forms must not be overlooked. Although the simplest techniques should be used, more sophisticated ones must be available when it is natural or necessary for the lawyer to use them. Where possible, sophisticated techniques should be made to appear simple to the lawyer builder.

In this context, we should say that our experience with production rules as a form of representation has been disappointing. We have found that lawyers do not find it easy or natural to formulate their knowledge as production rules.17 We are not saying here that production rules are not a suitable device for building expert systems; only that they do not appear to be a natural form for use by untrained lawyers.

2 End-users will be lawyers or para-legals: For some time to come, we believe that it is unlikely that legal expert systems will be used directly by those without legal knowledge. There are several reasons for this.

First, because of the difficulty of construction, expert systems will for some time be confined to relatively narrow specialist areas of the law. This means that people unfamiliar with the law may not even recognize that their problem falls within the area of expertise addressed by the system.

Secondly, the problems associated with evaluation of the performance of legal expert systems will make non-lawyers suspicious of their operation, at least if the legal problem is important enough that the user seeks legal advice in the first place. The user who is familiar with law and legal reasoning will be convinced by the justification which must be provided by the system.

Finally, there are probably problems of product liability which are best solved by keeping the distribution of systems confined to those who are able to exercise an independent judgment.

If this is correct, then once again, several corollaries flow from it. Again because of the relative level of computer illiteracy among lawyers, the user interface probably needs to be better than the existing standard for expert system. Generous "help" facilities must be available not only for the operation of the system itself, but for convenient probing in greater depth when unknown terms are encountered.

Another interface problem which is particularly acute in legal systems is the control of the depth of questioning. The problem may be illustrated by an example: in one of the DataLex applications program,18 it is necessary to know the place of formation of a contract. In the majority of cases, the user will know the answer if the question must be asked. In some cases, the user must reply "don't know", a reply which means that the system must probe to determine whether the contract was made within the jurisdiction. If the system asks questions which are "irrelevant" in the sense that they are directed to answering a question which has an answer which is obvious to the user, then the user will lose patience with the system.

The problem exists with all expert systems, but it is more pronounced in law because the depth of questioning in a situation such as that above must be determined not by parameters which characterize the general level of knowledge expected of users, which may be predicted, but by parameters which depend upon the particular problem at hand, parameters which may not be designed into the system.

There is a second corollary of this design criteria. In Australia at least, the overwhelming majority of practitioners are in small firms and in many cases in very remote areas. The only computing facilities available are likely to be the word processing system. While we did not consider it feasible to deliver expert systems on most word processors, we do believe it possible to deliver systems for IBM compatible personal computers and the Apple Macintosh.

The desire to deliver on these machines has led us to a somewhat unusual choice of language for an AI project. All DataLex modules are programmed in C, it being our perception that it is the most transportable of the computer languages.

3 Full Justification of Advice Must be Provided: Justification of the advice given in a legal expert system is as important as the advice itself, for if the justification includes reference to primary source material, as we believe it must, then the lawyer user will be in a position to form his or her own opinion of the worth of the advice received. The usual "why" and "how" facilities of production rule systems are simply inadequate for the problem of providing this level of justification.

If the justification is to be used to fullest effect, then the output should be in a form which is familiar to lawyers. This, in turn, will depend upon the particular use that the system is designed for, but in the case of advice giving systems of the sort that we are concerned with, the output is generally in the form of a barrister's opinion. Facilities are also provided which would allow for automatic document generation in suitable cases.

4 "Expert" is a relative notion: The literature contains many sentiments such as the following:

...expert systems in law ought to be designed to replicate legal experts, the knowledge represented, therefore, necessarily being of a depth, richness and complexity normally possessed by such a human being.19

As a standard at which to aim when constructing a system, we have no quarrel with this, but if it is meant that expert systems must achieve this standard before being useful, then we disagree. The measure must surely be one of cost effectiveness, particularly in those situations where the cost of human advice is prohibitive to the user.

It is not just the "impoverished client" situation where the costs of human advice is relevant. If success of an expert system is defined by the number of times that it is used, then one of the most successful known to us is a program that screens clients who think that they have a problem for the human expert. It turns out to be a simple matter to determine if the client does or does not have a problem which should receive the attention of the expert.20 This is a situation where previously the expert himself spent the time to conduct the screening. Such a highly specialized job would never have warranted the hiring of a human screener. The system works and is used although no one would ever suppose that it contained high level knowledge of the kind referred to above.

Susskind also identifies another way in which current expert system technology falls short of producing human capabilities:

...the crucial, and non-deductive, stage in legal reasoning is that of classifying the particular facts of the case, a stage in which none of the current computer systems are of support.21

Again, no one is likely to question the truth of the observation. However, most expert systems are aimed at capturing the knowledge of specialists so as to make that knowledge available to the general practitioner. In that context, the classification of facts is performed by the general practitioner regardless of whether it is a human or computer expert which is being consulted. Of course the specialist might say that the facts do not really fall into the specialist domain; a specialist expert system should have the ability to recognize problems which do not fall within its domain or which are borderline in some sense. The PANNDA system discussed below does just that.

LES: the DataLex Shell

The DataLex shell is named LES, an acronym derived from Legal Expert System. It currently has two fully integrated modules which allow the the builder the use of differing logics but which are transparent to the end-user. There are other operational modules which are not yet fully integrated into a single working environment, including an inductive decision tree generator and a free text retrieval module.

LES.DN - Decision Networks:

Our experience with production rules led to the observation that the lawyers who were attempting to define rules were actually drawing decision nets first and then writing the rules from them. It is, of course, obvious that every existing expert system could theoretically be rebuilt as a decision net but it is conventional wisdom that this is not a practical approach because of the "combinatorial explosion". We have found that the theoretical "explosion" was simply not occurring in the problems in which we were interested. The trees were highly pruned or there was a very high degree of duplication in the nodes or both. We investigated the possibility of providing machinery which would take the decision net as its starting point.

LES.DN allows the creation of decision networks in a modular fashion. Sub-nets may be entered at any point with return to the main net at the completion of the sub-task. This facility is also used as a part of the standard "help" mechanism. Typical LES applications have technical terms displayed in reverse video; the end-user may access the "help" sub-net for the term by typing the term followed by a question mark.

The modular construction of LES nets also facilitates program maintenance, one of the major problems of using structured nets instead of production rules. The modular approach also facilitates the control of question depth mentioned above.

LES.DN allows global attributes to have values which are available at any node in the net; some ordinary arithmetic facilities are also available so that, for example, the legal concept of a per stirpes distribution of assets upon an intestacy may be handled in a straightforward recursive fashion within the LES framework.

It is expected that LES.DN nets will provide the fundamental component of most applications. Although the knowledge base and the inference mechanism are not separated as in the production rule model, the system is, of course, still a knowledge based system, the inference mechanism consisting of the explicit interconnections between the nodes which are activated by responses supplied by the end user.

We see a number of advantages to using decision networks as the fundamental building block in a legal expert system. They include:

  1. * The above mentioned ease with which lawyers take to the model. This is probably because lawyers are well practiced in providing explicit reasoning even when the conclusion is not contentious;

  2. * Even in more complex problems, much of the relevant law may be quickly captured by the decision net. If a point is reached when the method is no longer appropriate, then other methods can interface relatively easily with the net;

  3. * The net allows a high degree of control over the operation and appearance of the system. As noted above, this is seen to be important since it is unlikely that lawyer users would have much tolerance for a poor user interface;22

  4. * Decision networks can by used to compile relatively sophisticated end reports and justifications by means of simple event driven mechanisms.

The actual application is written in a simple language which lawyers seem to find easy to learn. See appendix I for a sample from the application INTEST. The resulting text file is then compiled by the LES generator program into a C program file which is then itself compiled in the usual way to produce the final executable decision net.

2 LES.PANNDA: coping with case law

PANNDA, an acronym for "precedent analysis by nearest neighbour discriminant analysis", uses measures of similarity and statistical analysis to simulate the reasoning used by lawyers when dealing with case law.23 It was originally developed as an experimental technique applied to an area which is virtually 100% case law, the law which governs the right to possession of chattels which have been lost or abandoned.24

In a PANNDA application, the end-user is interrogated to determine the values of the legally significant parameters of the new problem on which advice is being sought. These significant parameters have been defined by the application author who has also defined the values of the attributes for the cases in the knowledge base. In the current implementation, the weight which is assigned to each of the attributes is determined statistically and calls for no input from the applications author.

Various statistical techniques are then used to determine which of the cases in the knowledge base is the "most persuasive" and most likely to be followed. There are currently two subsidiary checks which are designed to identify cases which are beyond the competence of the system and to give an appropriate warning to the end-user.25 The system also advises which of the cases are most likely to support a contrary conclusion and the final report is presented in a form which attempts to justify the conclusion reached as well as giving arguments which would "distinguish" the case which supports the opposite view. A sample report for a recent English Court of Appeal decision is shown in Appendix II.

Although originally designed as an independent system, the PANNDA module is now fully integrated with the DN module, making it possible to develop applications which require a combination of statute and case law.

It is also thought that the PANNDA module will be a valuable tool in solving the "discretion" problem. There are many situations where the end advice is that "the court has a discretion to....". It would be nice to be able to go further and give some arguments which would influence the court as well as provide an estimate of the likely outcome. The PANNDA approach supplies a powerful tool for tackling these sort of problems, although it seems likely that a completely successful system must call on the expert author to assign subjective weighting to the attributes which could be used in a probabilistic way to judge the reliability of the final advice.

3 AIRS: Free Text Retrieval

AIRS is an acronym for "Another Information Retrieval System".26 AIRS emulates the STATUS commands which are used in the commercial Australian CLIRS legal information data base. AIRS is intended to be used for teaching students the techniques of automated legal information retrieval and to be an integral part of the LES shell. Since free text retrieval is not often thought of as an integral part of an expert system, perhaps a few words explaining our views might be in order.

Given the importance of the primary materials in the legal domain, we believe that an integrated free text system can be valuable in several circumstances:

  1. * The primary materials may be freely accessed during the justification stage;

  2. * Until systems improve dramatically, there will always be complex predicates which must be left to the judgment of the end-user. As an example, the system may need to know if a party is "a fit and proper person". If the system itself is unable to provide any assistance in this matter, the text retrieval system at least allows the user to research the question in place, returning to the "normal" expert system when satisfied as to the nature of the answer;

  3. * Many technical terms are defined in statutes. Allowing direct access to those definitions achieves an economy at the time of writing the application since those terms need not be reproduced, but even more importantly, there is an automatic updating of the definitions in the event that the statutory definition changes.

AIRS is not yet fully integrated into the LES package.

4 Other inferencing modules

We have already indicated our reasons for beginning development of the LES shell with a decision network module and a case analysis module as the primary components, but we would like to emphasize that one of the main purposes of the shell is to incorporate a number of useful legal inferencing procedures within a consistent environment so as to maximize the choice of tools available to the applications author. The following modules are under development:

Inductive decision tree generator: this procedure is based on an algorithm which is employed in the ID3 system.27 The technique "automatically" generates a decision tree from a set of "examples". In this sense, the system is "inductive", forming a set of rules from a collection of examples.

Early experiments with the inductive tree generator suggest that it is a valuable tool for producing trees which may then be "hand crafted" into a more precise and detailed LES decision network. In addition, the method should be very useful in areas where formal legal reasoning is less apparent but where there is a plethora of examples. Such areas may include the size of awards for personal injuries and sentencing practices.

The inductive decision tree generator is not yet integrated into the LES shell.

Baysian inference: the existing LES shell has no procedures for reasoning under uncertainty, yet such a facility seems essential if the "discretion problem" is to be handled adequately by an expert system. We believe that this might be best accomplished by the incorporation of Baysian methods, since it seems likely that the lawyer/author could reasonably be expected to place numbers resembling prior probabilities on various situations. Program specifications have been drawn, but much experimentation remains to be done.

Production rules: although our experience suggested that production rule models were not the most logical starting point for a legal expert system shell, there is no doubt that they represent a suitable method for handling some parts of a legal problem. We anticipate development of a production rule module in the LES environment.

5 Existing applications:

During 1986, the LES shell has been made available to a number of interested lawyers who have commenced work on applications. Perhaps the most interesting characteristic of this work is that the applications authors have, for the most part, no experience in computing. In one case, a group of students from the University of Sydney Law School brought an application to the demonstration stage within six weeks. The following is a brief description of some of the more advanced applications.

INTEST

advises on the distribution of property on intestacy in New South Wales.28 The package provides complete coverage of an area which is almost entirely statutory, but which includes relatively complex numerical calculation if the family tree is complex.

COPYRITA

advises on the law of copyright in Australia.29 At the present time, advice is given on whether the subject matter is protected by copyright and whether there has been an infringement of that copyright.

FINDER

uses the PANNDA module to advise on rights to possession of chattels which are lost or abandoned. This is an area which is almost entirely case law. The body of case law is notorious for its lack of consistency and were said to be the "delight of professors and text writers, whose task it is to reconcile the irreconcilable".30 A sample FINDER opinion is attached as Appendix II.

Others

Several applications are being written by groups of students at the three universities represented by the authors of this paper. These applications include programs which give advice on the availability of service outside the jurisdiction (LONGARM: University of Sydney), the notional estate entitlements defined by the Family Provisions Act 1982 (NOTATE: University of New South Wales) and entitlement to workers' compensation payments (COMPO: NSW Institute of Technology).

Acknowledgments

The authors gratefully acknowledge support for the DataLex project from the Law Foundation of New South Wales, the Australian Research Grants Scheme, the University of New South Wales Faculty of Law, the New South Wales Institute of Technology and the University of Sydney.

1 See, for example, Hayes-Roth, F, Waterman, D A and Lenat D B (eds) Building Expert Systems, Addison-Wesley, 1983.

2 Michie and Johnston, The Creative Computer, Pelican, 1984

3 Sergot, M J, et al, "The British Nationality Act as a Logic Program" 29(5) Communications of the ACM

4 Commissioner of Inland Revenue v West-Walker [1954] NZLR 191. The statute is section 163 of the Lnad and Income Tax Act 1923 as re-enacted by s.12, Finance Act (No 2) 1948.

5 Waterman, DA, Paul, J and Peterson, M "Expert Systems for Legal Decision Making" in Proceedings of the Second Australian Conference on Applications of Expert Systems, NSW Institute of Technology, 1986

6 Waterman, DA, Paul, J and Peterson, M "Expert Systems for Legal Decision Making" in Proceedings of the Second Australian Conference on Applications of Expert Systems, NSW Institute of Technology, 1986

7 Indeed, the standard rules are entirely contradictory, a fact which is acknowledged in the aphorism that rules of statutory interpretation "hunt in pairs".

8 Pearce, DC Statutory Interpretation in Australia (2nd Ed, Butterworths.

9 Stone, J Precedent and Law, 1985, Butterworths, Sydney.

10 Of course, the idea of intelligent pre-processors is not new. See Bing J Handbook of Legal Information Retrieval, 1984, Elsevier, Amsterdam. And some of the most important expert systems work has been directed toward "intelligent" information retrieval; see Hafner, C An Information Retrieval System Based on a Computer Model of Legal Knowledge, UMI Research Press.

11 (1929) 34 Com Cas 263.

12 at p273.

13 "...the best drafted Act of Parliament ever passed", Bank Polski v K J Mulder & Co [1942] 1 KB 497, 500, per MacKinnon LJ.

14 McCarty, T "Intelligent Legal Information Systems: Problems and Prospects" in Campbell, C (ed) Data Processing and the Law, Sweet and Maxwell, 1984, London.

15 Susskind, R E "Expert Systems in Law: a jurisprudential approach to artificial intelligence and legal reasoning" [1986] Mod L Rev 168, 185-6.

16 Hayes-Roth F, Waterman D A, and Lenat D B Building Expert Systems, Addison-Wesley, 1983.

17 This finding is, of course, not inconsistent with the experiences in other fields. It is the essential reason for the existence of the knowledge engineer.

18 LONGARM, a system which advises on the availability of service of originating process outside the jurisdiction of New South Wales. The system was built by students at the University of Sydney.

19 Susskind, R E "Expert Systems in Law: a jurisprudential approach to artificial intelligence and legal reasoning" [1986] Mod L Rev 168, 185.

20 The system determines if the client actually has a problem which falls within the jurisdiction of a part of the Trade Practices Act 1974. It was built by Sydney solicitor Philip Argy who is the expert whose time was previously being used.

21 Susskind, R E "Expert Systems in Law: a jurisprudential approach to artificial intelligence and legal reasoning" [1986] Mod L Rev 168, 190.

22 Of course, it is possible to provide that control with a production rule model, but the methods are not easy or natural. Nor, it might be said, are they in conformity with the spirit of the production rule model.

23 See Cover, T M and Hart, P E "Nearest Neighbour Pattern Classification" IEEE Trans Inform Theory Vol IT-13, p21.

24 Tyree, "The Geometry of Case Law" (1977) 8 VUW L Rev 403; "Finders Keepers: a quantitative analysis of 'finders' cases" in Brook, et al (eds) The Fascination of Statistics, 1986, Marcel Dekker, New York, pp73-88; "Will Justice Fall to Bits? Expert Systems in Law", (1986) 62 Current Affairs Bulletin 13-18

25 The primary classification technique is the nearest neighbour algorithm. The subsidiary methods concern classification according to nearest centroids of the sets of cases in the data base and a final check that the selected cases are not too near the boundaries.

26 Written by Andrew Mowbray. The information here is from Mowbray, "A AIRS Free Text Retrieval System (Pre-release information)" DATALEX Project, New South Wales Institute of Technology, Faculty of Law. See also Greenleaf, Mowbray and Tyree "Legal Expert Systems: Words, Words, Words...?", paper presented at the 1st Australian Artificial Intelligence Congress, Melbourne, 18-21 November 1986.

27 See Quinlan, J R, et al "Inductive Knowledge Acquisition: A Case Study", in Proceedings of the Second Australian Conference on Applications of Expert Systems, NSW Institute of Technology

28 INTEST was written by Andrew Mowbray and Rosalind Atherton of the Faculty of Law, University of New South Wales.

29 COPYRITA is being developed by Graham Greenleaf and Phillip Griffith, Senior Lecturer in Law, New South Wales Institute of Technology with research assistance from Karen Lever B Juris, LL B (WA).

30 per Lord Goddard, Hibbert v McKiernan [1948] 2 KB 142,149