Alan L Tyree

Generating Legal Arguments 1

The Legal Background

The legal system of most former British colonies, including that of the USA, is based on the English system of Common Law.  These legal systems have each evolved in different ways so that the detailed law on any subject might be different, but the Common Law uses reasoning methods that constrain and direct the directions that legal development might take.  These reasoning methods appear to be unique to law and pose some interesting problems for the application of artificial intelligence to legal problems.

The most characteristic feature of the Common Law is the doctrine of precedent.  According to this doctrine, each and every case decided by the court becomes a part of the law itself, not merely an explanation of the law which later judges may adopt or reject.1

The danger in such an doctrine is that the law may become completely inflexible.  In order to prevent this, there have been a number of devices developed to avoid "unwanted" outcomes.  These devices include the obvious ones such as having regard to the seniority of the court which decides the case, but also less obvious ones which become a part of the legal reasoning process.

The process of reasoning with case law is so fundamental to the development of the Common Law that it forms a major part of the curriculum in law schools.  Students are taught to read a body of case law, to apply that law to new fact situations and to use the legal technique of "distinguishing" previous cases.

This form of legal reasoning is in some respects similar to the general problem of reasoning by analogy, but there are some specific and difficult differences.  First, there is a deontic aspect, for even those cases which are not strictly binding on a court have some legal force.  A second difference is that the process contains formalised methods for "distinguishing" past cases.  A case is distinguished by noting legally significant differences from previous cases which may be used to justify rejecting the outcome of the previous case insofar as its applicability to the problem at hand is concerned.

At one time it was thought that each decided case stood for a single rule of law which could be formulated in much the same way as a section of a statute, but this theory has been shown to be completely unworkable.2  It is now clear that any interpretation of the legal significance of a case must be in the larger context of the legal material in which it is embedded.3  The meaning of a case can and does change as the body of knowledge surrounding it changes.

Perhaps for this reason, there has been little success in dealing with the general problem of reasoning with case law by means of the usual production rule formulation.  It is not that it is theoretically impossible to write such rules, but that it is not the natural way in which lawyers reason with cases.  If this is correct, then attempting to use the usual rule based systems will probably result in mediocre performance of the system and will certainly be wasteful of time in the knowledge engineering process.

Statute Law

It might be thought that systems which deal only with statute based law might avoid this problem of specialised common law reasoning, but that is not the case.  This is because most statutes have received interpretation by the Courts which either clarify or change their meaning.  Furthermore, the interpretation of statutes is itself a reasoning process which is unique to law.

Every lawyer has a favourite example of a statute which does not mean what it would appear (to a non-lawyer) to say.  Here is ours:  the statute reads "...every person shall..".4  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...)".

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.

The result was reached by applying well known and well understood techniques of legal reasoning.  Some of those techniques involve the application of old case law; others are rules of interpretation used by lawyers when reading statutes.  However, these rules of statutory interpretation are not in a form which allows them to be expressed directly in a production rule system.5  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.6

It is clearly possible to write rule based systems which would assist in applying that reasoning to assist in the interpretation of statutes, but the point is that a simple translation of a statute into executable code is likely to produce a machine lawyer that is a candidate ripe for disbarment.

In spite of these difficulties, there have been some useful rule based legal systems developed.7  Although most of these deal with statute law, there are a few which have attempted to deal with case law related subjects.8  For the reasons mentioned above, we believe that an alternative approach to the problems of case law will produce better systems at a cheaper cost.

Example: the finder cases

As part of the DataLex Project, one of us has explored an alternative which makes use of concepts of distance between cases.9  The resulting program is called, for reasons which will become clear, FINDER.10

Although every serious legal problem will include both statute and case law reasoning, there are a few areas which have been untouched by statute.  The law in these areas is to be extracted entirely from the decided cases.  These areas are clearly attractive for experimentation purposes.

One such area is the dispute which arises when a person finds an object which is then claimed by another party.  In the proto-typical case, the item is found on premises which are occupied by the other claimant.

The law which governs such disputes is contained almost entirely in the decided cases.11  Although there are not a large number of cases, they are complicated and difficult to resolve.  Since each of the parties is seeking a windfall, policy is difficult to formulate and there is little sense of justice to guide the court.  Of these cases, one judge has said:12

These cases...have long been the delight of professors and text writers, whose task it often is to attempt to reconcile the irreconcileable.

Although the cases go back to 1721, the most recent one is Parker v British Airways.13  Parker was a passenger on British Airways.  While waiting in the VIP lounge, he found a gold bracelet lying on the floor.  He took the bracelet to lost and found and left it with instructions that it was to be returned to him if the owner of it was not found.  Although BOA did not find the owner, they did not return it to Parker but instead sold it for some £800.

This paper describes the FINDER system which was designed as an experimental testbed using the finder cases.  Parkers case is not included in the knowledge base of the FINDER system.  An opinion written by the system is included as appendix A.

Legal Expert Systems

The form of the output of FINDER illustrates an important feature of legal expert systems.  Justification plays a different role in a legal expert system than in other advising systems.  Usually the justification mechanisms are rudimentary and are clearly designed to give the user confidence in the decision reached by the system.  A system may advise "take two aspirin and go to bed"; the justification provided for that advice will usually consist of a listing of the rules used to derive the conclusion.

By contrast, the justification in a legal system is the main product, for the justification is no more and no less than the legal arguments which support the suggested outcome.  It is in the nature of legal reasoning that these arguments must also address the support for the opposite outcome.  It is these arguments which, if the matter goes to court, must be presented for adjudication.

In this regard, a legal expert system is more akin to design systems than to the traditional diagnostic systems of medicine.  The FINDER system uses a simple but effective model for the construction of legal arguments and presents them in a form which is familiar to lawyers.

The PANNDA system

The reasoning model used in the FINDER system has been incorporated as a separate reasoning model in a more general legal expert system shell called XSH.14  XSH supports the usual procedural and rule based reasoning, but also allows for case-based legal reasoning on the FINDER model.  The module is called PANNDA, an acronym for Precedent Analysis by Nearest Neighbour Discriminant Analysis.

Knowledge Representation

Cases are represented in PANNDA by means of frames.  In the FINDER implementation, the frames are "flat", ie, they are vectors.  The vector slots represent the legal facts of the case which were determined by experts in the area to be of legal significance.

"Legal facts" may contain a significant amount of expert knowledge themselves.  Thus, one of the facts of interest in FINDER is that one of the parties was, or was not, the occupier of the premises.  "Occupier" is itself a term of some complexity, so that it would be possible, or at least useful, to construct an expert system adviser to identify if a person is or is not an "occupier" of premises.

In FINDER, this representation of the cases is the sole input of expert knowledge into the system.  In our experience, lawyers are able to identify the slots and to use them to represent cases in a number of areas without a need for close supervision by a "knowledge engineer".  This self-sufficiency is an important design criterion in all of the work done by the DataLex Project, for we believe that it is a mistake to distance the expert from the final product.

Inference mechanisms

In a consultation, FINDER solicits information from the user which will allow the construction of the full case information for the user's situation.  At the current state of development, there are no checks that the user really does have a finder problem, although there are several consistency checks later in the process which may serve to identify that problem.

Once the information from the user has been obtained, FINDER calculates a measure of similarity between the client user's case and all other cases in the knowledge base.

Although a number of different algorithms have been used in FINDER, the current one is the original and performs as well as any.  A weighted Euclidean distance is used as the basic measure of dissimilarity.  The weights could be supplied by the expert but, if not, they are generated automatically by weighting each variable an amount inversely proportional to the variance of that variable across the knowledge base.

This "inverse variance" weighting requires a few words of explanation, since ordinarily high variance variables are more important in discriminant analysis.  Our reasoning was that a variable is included in the knowledge base because the expert believed that it was legally significant.  If the variable has a very low variance, then it is not significant for its discriminating powers; we reasoned that it must be because of its greater legal significance.15

Nearest Neighbour

Once the distances are calculated, FINDER searches the knowledge base for that case which is "closest" to the user described case.  The system then predicts that the outcome of the users case will be the same as that of the "nearest neighbour".

The use of nearest neighbour discriminant analysis in preference to some other form is to some extent arbitrary.  There are some attractive theoretical properties of nearest neighbour analysis when the underlying statistical distribution of the variables is unknown and the connection of the method with the minimal spanning tree provides some promising lines of inquiry.16

As a practical computing matter, nearest neighbour methods are appealing because the computational time increases only linearly with the size of the knowledge base.  This is in marked contrast to the well known "combinatorial explosion" effect when using rule based systems.

Nearest Other

The nearest neighbour is the case which lends most support to the predicted outcome and it is the case which is used to construct the argument for the predicted outcome.  However, it is a foolish lawyer who chooses to prepare only one side of a case.  As in all argument, one of the most powerful methods of preparation is to prepare the argument for the opposing side.

In FINDER, this takes the form of finding the case which most strongly supports the conclusion which is opposite to that of the predicted outcome.  The knowledge base is split into two parts and the cases with the same outcome as the nearest neighbour are excluded from consideration.  The remaining cases are searched for the nearest neighbour.  When found, it is call the "nearest other" by the FINDER system.

Anomalous cases

In our opinion, all legal expert systems should be able to recognize cases which are "difficult" in some sense and refer those cases to a human expert for consideration.  In FINDER, this facility takes the form of several cross checks for consistency.

The first check is to use a second form of discriminant analysis to rework the problem.  Since nearest neighbour analysis is prone to problems of "chaining", it seems sensible to test it against a method which is more "globular" in outlook.  One simple globular method is to compute the centroids of the two major groups of cases in the knowledge base and then measure the distance from the client user's case to the centroids.

A second consistency check is also designed to catch examples of "chaining" in the nearest neighbour model.  This check uses the nearest neighbour method itself to detect anomalies.  Although the basic system makes decisions on the basis of a single nearest neighbour, the method easily extends to the k-nearest neighbour approach.  If the ordered list of nearest neighbours contains cases which are "too close" and which contradict the predicted outcome, then there is cause for some concern.

The REPORT System

The REPORT subsystem is relatively simple, but it produces legal opinions which are in a form which is recognisable to lawyers.  The generated opinion follows a rigid format which may be described schematically as:

Announce the favoured outcome

Summarize the case which most strongly supports the favoured outcome;

Summarize the most important features which the user's case has in common with the retrieved case;

Summarize the most important features for which the user's case differs from the retrieved case (these will be the features which the "other side" will try to use to discredit, ie, to "distinguish", the predicted outcome);

State the outcome of the case which most strongly favours the result opposite to that predicted;

Give a summary of the "nearest other" case;

Indicate the most important features in which the user's case differs from the retrieved "nearest other" case, ie, "distinguish" the nearest other;

State a conclusion.

Items 2 and 4 merely retrieve "canned" summaries.  The other steps in making the report rely upon a careful piecing together of phrases which are stored as translations of the "facts" which are used in the knowledge base.  An examination of the Parker opinion will disclose that there is some detailed attention paid to verb forms in the different parts of the opinion, but the basic structure of items 3,4 and 7 is the same.

Only the most important, ie, the most heavily weighted, factors are used in the written opinion.  In earlier versions it was found that including all factors made the opinion unreadable while adding nothing valuable to the legal argument.

Anomalous Case Reports

Anomalous cases, that is, user cases which do not pass the internal consistency checks, are treated in the same way as an ordinary case except that a warning is issued to the user that something is wrong.  The user is given the opportunity, indeed the advice, to abort the program and see a human adviser.  If the user insists upon continuing, the opinion is written in the usual way.  In my experience, most anomalous cases result in very bad opinions.

PANNDA and rule based systems: The policy problem

The law relating to finders is entirely case law.  Indeed, the area was chosen for experimentation for precisely that reason.  It is an unusual area in this respect, for almost all serious legal problems involve both statute and case law.

There does not seem to be any direct method of using the PANNDA approach for the analysis of statute law except for the somewhat artificial method of concocting hypothetical cases.  If this approach were used, it would be necessary to change the REPORT subsystem very substantially, for the current format does not really make sense for hypothetical cases.  Perhaps the case summaries could be replaced by some sort of argument structure which would reveal why the hypothetical case was chosen and why it is thought that the outcome is that suggested.

It seems more logical to handle "pure" statute law by means of rule based systems.  But even in the FINDER model, integration with a rule based system would make a great deal of sense.  One immediate suggestion is a "screening" rule which would take care of the problem of a finder who is a trespasser on the land of the other party.  In such a circumstance, there is scarcely any reason to continue with the other questions or to invoke the FINDER inferencing model.

This is just one example of what lawyers call the "policy" issues involved in finders cases.  There are certain circumstances which have never arisen in reported cases for the simple reason that the answer is so obvious that only a fool would attempt to defend the issue in court.  

Another example might be that the place where the object is found is a bank vault;  although the finder may not be a trespasser, the occupier of the premises, the bank, clearly permits the entry only for very limited purposes and it would scarcely make any sense to suppose that the finder of an article in such a place would be entitled to it in preference to the occupier.

Although lawyers refer to these as policy issues, they are actually rather closely tied to fact situations which can be described in the knowledge base language.  Again, these situations could be the basis of hypothetical "policy" cases in the knowledge base.  The cases could then be given a very high weighting in the nearest neighbour algorithm to ensure that "policy" was not lightly overridden.

However, it seems more straightforward to phrase these policy situations as rules.  If that is done, then the problem becomes one of devising a smooth interface between the rule based knowledge system and the PANNDA model.

One way of doing this has been implemented in XSH.17  In essence, there is a single rule which initiates the session:


The initiating goal is to find a value for OUTCOME.  Determination of POLICY is done by means of backward chaining production rules.  Once the determination of the value of POLICY is complete the system proceeds to the PANNDA module to determine the case law aspects of the problem.

This approach has worked well in the finders cases, eliminating the asking of irrelevant questions when the situation is one which falls within the policy considerations of the finders cases.  The rule based procedures also manage to filter out some anomalous cases which would otherwise need to be handled by the consistency checks in the PANNDA module.

The mixed model also provides another valuable consistency check, for if the POLICY and PRECEDENT variables indicate different outcomes, then the case is clearly "difficult" for the system.

Further Development

In the real Parker case, a great deal of the judgment of the English Court of Appeal was devoted to a discussion of whether Bridges v Hawkesworth was wrongly decided.18  It would be nice if FINDER could be developed to recognize those cases which may be candidates for being overruled and to develop the arguments for such a position.

One possibility is to delete the case from the knowledge base and then to treat it as a new user case.  If this is done with Bridges v Hawkesworth, then the FINDER system considers it to be an anomalous case.  If pressed for an opinion, FINDER considers that the outcome in Bridges v Hawkesworth is that the finder should win, ie, the same as the real Bridges case and the same as found by the Court of Appeal.

Although the results are correct in the sense of coming to the right conclusion, they are scarcely satisfactory in the overall result.  The system upholds Bridges v Hawkesworth on the basis of a later case, Hannah v Peel.19  The problem is that Hannah v Peel was itself founded on Bridges v Hawkesworth, so the opinion has the look and feel of a very biased adviser.

The procedure presupposes that the user wishes to question the validity of Bridges v Hawkesworth.  That is not an unreasonable assumption, since the only reason to question the validity of Bridges, or any other case, is that it appears to support an outcome that is unfavourable to the user.  In this model, the system is acting more as a research assistant than as an adviser.


The authors gratefully acknowledge the assistance of the Australian Research Grants Scheme, the Law Foundation of New South Wales, and the Faculties of Law of the University of Technology,Sydney, the University of New South Wales and the University of Sydney.

Appendix A

Sample Opinion Written by FINDER program

The following is an opinion written by FINDER given the facts of Parker v British Airways [1982] 1 All E R 834.  Parker found a gold bracelet in the public area of the VIP Lounge of British Airways.  He turned it in to lost and found with instructions that BA should attempt to find the true owner of the bracelet.  Failing that, Parker asked that the bracelet be returned to him.  BA sold the bracelet when they were unable to find the true owner.  BA leased the terminal where the lounge was located.

The consultation begins by the system soliciting relevant facts from the user.  When sufficient information has been gathered, the system writes an advice.  The opinion below is presented exactly as it comes from the computer.


In my opinion, the outcome is the same as that in Bridges v. Hawkesworth, that is, the finder should win.

In Bridges v. Hawkesworth (1851) 21 LJQB 75 the plaintiff found a bundle of banknotes on the floor of the public area of a shop.  He handed the notes to the shopkeeper in order that the true owner of the notes might be found.  Although the owner never was found, the shopkeeper refused to return the notes to the finder.  The court found for the finder, holding that there is a "general right of [a] finder to any article which has been lost as against all the world except the true owner".  It was further noted that the notes had never been in the custody of the defendant nor within the protection of his house as might be the case had they intentionally been deposited there.

There are many similarities: the finder is not the occupier, the chattel is not attached, there is a bailment of the chattel and the chattel is not hidden.

Of course, the present case is not on all fours with Bridges v. Hawkesworth since in that case the non-finder was the owner of the real estate.

The opposite result was reached in Yorkwin v. Appleyard.

In Corporation of London v. Appleyard [1963] 1 WLR 982 workmen employed by Wates Ltd were engaged in cutting a key-way into a cellar wall for the purposes of securing a foundation when they found an old wall-safe built into a recess of the old wall.  Inside was a wooden box which contained a large number of Bank of England notes.  The notes were handed over to the City of London police who sought interpleader proceedings to determine who was entitled to the possession of the notes.

Wates Ltd were an independent contractor engaged by Yorkwin Ltd for a construction project.  Yorkwin were lessees in possession of the property which was owned in fee simple by the Corporation of London.

The Court followed the decision in South Staffordshire Water Co v. Sharman [1896] 2 QB 44 in holding that the occupier is, in the absence of a better title elsewhere, entitled to the possession of objects which are attached to or under the land.  Consequently, since the notes were in a wooden box within a safe built into the wall of the old building, the safe formed part of the demised premises.  Yorkwin, being in lawful possession of the premises, were in de facto possession of the safe, even though ignorant of its existence.

Although Yorkwin was entitled to possession as against the finders, they in turn were displaced by the Corporation of London which relied successfully on a term in the lease which granted them the right to certain objects found on the premises.

But Yorkwin v. Appleyard has some significant differences: the chattel was attached, there was not a bailment of the chattel, there was a master/servant relationship between the parties and the chattel was hidden.  Consequently, I think there is nothing in Yorkwin v. Appleyard to warrant any change in my conclusions.

1  (1989) 2 Journal of Knowledge Based Systems 46.

1  This is different from most European legal systems.  In these Civil Law systems, a judge of a lower court is theoretically free to differ from the judgment of even the highest court in the land.  Chaos is averted by a system of judicial review in which senior judges determine the promotion of more junior judges.

2  Goodhart, Essays in Jurisprudence and the Common Law, Cambridge; The Ratio Decidendi of a Case, (1959) 22 Modern Law Review 117; Stone . The Ratio of the Ratio Decidendi (1959) 22 Modern Law Review 597.

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

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

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

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

7  See, for example, 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 and the references therein.

8  See, for example, Rissland and Ashley, "A Case-Based System for Trade Secrets Law"; Smith and Deedman, "The Application of Expert Systems Technology to Case-Based Law", both in The Proceedings of The First International Conference on Artificial Intelligence and Law, May 1987, Boston, Mass.

9  The DataLex Project is a joint research project into legal expert systems.  Responsibility for and management of the project is shared among the three authors.  Alan Tyree is responsible for the work on FINDER.

10  For other background on the FINDER system, see Tyree, The Geometry of Case Law, (1977) 4 Vict U L Rev 403;  Fact Content Analysis of Case Law, Methods and Limitations, 22 Jurimetrics Journal 1; Greenleaf, Mowbray and Tyree, "Expert Systems in Law: the DataLex Project" in The Proceedings of The First International Conference on Artificial Intelligence and Law, May 1987, Boston, Mass; Tyree, Greenleaf and Mowbray, "Legal Reasoning: the problem of Precedent" in Proceedings of AI'87, Sydney, 1987; Tyree "Finders Keepers: a quantitative analysis of 'finders' cases" in Brook, et al (eds) The Fascination of Statistics, 1986, Marcel Dekker, New York, pp73-88

11  Various jurisdictions have legislation for particular kinds of objects, eg, explosive materials or native artifacts.  Such special objects should be handled by a separate rule based module.  See below.

12   Hibbert v McKiernan [1948] 2 KB 142, per Lord Goddard.

13  Armory v Delamirie (1721) 1 Strange 505; Parker v British Airways [1982] 1 All ER 834.

14  XSH was written by Andrew Mowbray as part of the DataLex Project.

15  This suggests that some direct measure of information might be a better method of weighting.

16  See Cover, T M and Hart, P E "Nearest Neighbour Pattern Classification" IEEE Trans Inform Theory Vol IT-13, p21 for a discussion of the statistical properties of the nearest neighbour algorithm.  For the connection with the minimal spanning tree of the complete graph, see Tyree, Greenleaf and Mowbray, "Legal Reasoning: the problem of Precedent" in Proceedings of AI'87, Sydney, 1987.

17   Written by Andrew Mowbray of the University of Technology, Sydney, as part of the DataLex Project.

18  (1851) 21 LJQB 75.

19  [1945] 1 KB 509.