Project report: Split-Up ? A Legal Expert System which determines property division upon divorce

Artificial Intelligence and Law 3: 267-275, 1996. © 1996 KluwerAcademic Publishers. Printed in the Netherlands.

Project Report: Split-Up - A Legal Expert System which Determines Property Division upon Divorce J O H N Z E L E Z N I K O W , A N D R E W S T R A N I E R I AND M A R K GAWLER Collaborative Law and Artificial Intelligence Research Project, Database Research Laboratory, Applied Computing Research Institute, La Trobe University, Bundoora, Victoria, Australia 3083 (Received 7 February, 1995; accepted 23 March, 1995) Key words: Legal Expert Systems, sequenced transition networks, neural networks, I]93 algorithm, Toulmin Argument Structures, case-based reasoning, production rule expert system, divorce, property division, explanation

1. Introduction Case based reasoning is the process of using previous experience to analyse or solve a new problem, explain why previous experiences are or are not similar to the present problem and adapting past solutions to meet the requirements of the present problem. Despite the fact that English common law places a high emphasis on the use of cases to resolve legal disputes, it is only in the last decade that legal knowledge based systems have directly used case based reasoning, although it is true that [Meldman 1977], [19lcCarty 1983] and [Smith and Deedman 1987] used cases to determine heuristics in rule based expert systems. As [Zeleznikow and Hunter 1994] state, many current commercially viable legal knowledge based systems must be essentially rule based. Though other technologies offer great promise, they are not sufficiently developed to be useful in building large, efficient, robust systems. One of the major early uses of case based reasoning was in the domain of dispute resolution [Kolodner and Simpson 1989] and [Sycara 1990]. Noting as [Ross 1980] states: The principal institution of the law is not trial: it is settlement out of court, led us to ask the question as to whether one can predetermine which potential cases will and will not be litigated. One reason for settling a legal dispute is that the litigants wish their negotiated agreement to remain secret and not act as a precedent which can be used by other parties. In the domain of Australian family law all cases, whether litigated or negotiated, are filed with the court, and hence available for perusal. This led us, in November 1991, to approach Graham, J. of the Family Court of Australia, with a view to using case based reasoning to analyse the differences between negotiated and settled cases. Graham, J. convinced us that our proposal was not realistic, because the reasons parties arrive at consent orders is often that they wish to keep certain facts secret, such as hidden assets and undeclared income. He however encouraged us to build a knowledge based system to model that part of the Family Law Act (1975) which deals with the distribution of assets upon the dissolution of a marriage. He claimed that whilst that part of the Act which deals with child welfare issues is discretionary, there are established mechanisms

268

JOHN ZELEZNIKOWET AL.

for how marital property is distributed. Graham, J. further claimed that the development of a system, to be used by quasi judicial court officials (registrars) amongst others, would encourage potential litigants to negotiate. The system which has been developed for this task is known as Split-Up. Split-Up end users are the parties to a dissolution, solicitors or mediators. Some end users are expected to have little, if any family law expertise and no general legal expertise. The system is only useful if it is able to be used by people with typical assets. This application is therefore quite different from many other expert systems where the expert system is required more for atypical cases and the end user is a semi-skilled expert.

2. The Split-Up System In determining the distribution of property under the Family Law Act (1975) a judge performs the following functions: 1. She determines assets of the marriagethe Court is empoweredto distribute. This task is denoted as the commonpool determination; 2. She determineswhat percentageof the commonpool each party to the marriagereceives,known as the percentage splitdetermination; 3. She determinesa finalproperty~order in line with decisions made in 1 and 2. Our approach to developing a system to provide advice about property distribution upon divorce was to use prototyping. Prototyping is the process of building an experimental system quickly and inexpensively for demonstration and evaluation so that users can better determine their information requirements. The prototype is a working version of part of (possibly all of) a preliminary model of the information system. Once it is operational, the prototype is further refined until it meets the user's requirement. Once the design has been finalised, the prototype can be converted into a fully functional system. We had been involved in prototyping a legal expert system CAAS (Credit Act Advisory System) for a firm of Melbourne solicitors, Allan Moore and Co [Vossos et al 1993]. Initially, a prototyping approach was taken to development, allowing intended users to participate in the design of the system while at the same time pleasing management by developing an initial model of the system which they were able to monitor. The prototype was designed using Neuron Data's Nexpert Object Development Environment: an interactive interpretive environment that makes use of graphical editors and browsers. Once the prototype had the approval of management, it was converted to a generic C++ program running under Microsoft Windows 3.1. The Split-Up prototype was a production rule expert system built using the VP-expert expert system shell [Stranieri and Zeleznikow 1992]. The system was able to determine which assets the court is empowered to distribute and the entitlements of each partner to the marriage. VP-expert is a rule based expert system shell for the IBM-PC. The principle inferencing mechanism is backward chaining, though it does allow a primitive form of forward chaining. Our Split-Up prototype incorporated certainty factors. It had three hundred rules in its rule base. The following factors were observed in the construction of the Split-Up prototype:

A LEGAL EXPERT SYSTEM

269

- The task of determining the common pool was suited to modelling using rule based systems; - The task of determining what percentage of the common pool each party to the marriage receives should not be modelled using rule based systems; - The knowledge acquisition task which was conducted by interviewing a domain expert is very time consuming.

In the following sections we discuss how we overcame some of the above-mentioned difficulties.

3. The common pool determination in Split-Up In considering the assets of partners to a divorce, a judge of the Family Court of Australia may reach one of three mutually exclusive conclusions: - The asset is excluded from the common pool and takes no part in property calculations; or - The asset is included in the common pool and the Court has the power to transfer the asset; or - The asset is included in the common pool but the Court does not have the power to transfer the asset.

The common pool determination task was suited to a rule based reasoning approach although the reasoning process for this task is not explicitly governed by any statue. Experience with the prototype revealed that heuristics relevant for the common pool deterruination are quite procedttral. Leading cases which resolve a question of uncertainty occur rarely and only a minority of litigated cases focus on the inclusion or exclusion of particular assets. The elicitation of expert heuristics for a rule based system was initiated using a structured interview technique with a domain expert, Renata Alexander, who has over twenty years experience with the Legal Aid Commission of Victoria. 1 The creation of if-then rules from transcripts of structured interviews proved to be a time consuming and cumbersome process. In order to attempt to accelerate this process, we encouraged the expert to represent her own dialogue with a hypothetical client as a directed graph. In this way, fifty one graphs containing two hundred and thirty nodes were elicited in thirteen, one hour sessions. Whilst we do not claim the use of directed graphs is suitable in all domains, we believe they are useful for the common pool determination, since this task does not involve many vague or ill-defined terms and the procedural steps used are not controversial. When the graphs were shown to a leading Australian family law academic, Richard Ingleby, he claimed they were both accurate and informative. Rather than convert the graphs to rule sets for use in a rule based expert system shell, we created a program which converts directed graphs into an inferencing system capable of forward and backward chaining and the provision of explanation. The methodology for converting the graphs is based on set theoretic definitions we called sequenced transition networks. A detailed description of sequenced transition networks can be found in [Stranieri et al. 1994a] though essential features are outlined below. I The Legal Aid Commission of Victoria is a government funded organisation which specialises in legal advice and representation for low income clients.

270

JOHNZELEZNIKOWETAL.

Figure 1 is an example of a graph which determines whether or not a vehicle will be included in the common pool. Arcs represent alternate client responses to the expert's queries. A program has been developed which provides the graph drawing facilities, numerically labels arcs and nodes and stores all text in a relational database. 2 Nodes are labelled with identifiers which correspond to paths taken to reach the node (illustrated in o u t ~ e ) . Thus the conclusion 'The vehicle will be excluded from the pool' can be reached by traversing path [0~21] or [0221]. During rtm time an independent program performs operations on sets of these paths to implement forward and backward chaining and create three types of explanations. 3 Rules for a conventional inference engine do not need to be written. All user prompts, conclusions and explanations are generated directly from text in nodes and on arcs of the graph. Experts can thus develop and maintain knowledge bases without learning the syntax of an expert system shell language.

0.

Initial node

2. a gift or "2 U inheritance ~ ~...,~Tbe gift " ~ 1. neithe~ a gift nor "k~ benefited ~] inher~tnee I ~.l~n,~-,.-, . ~ ' j . ~ : . " ~ !2. both •~t t av....vvHl~t¢ ~~

01/f~. k

-'~

! [

Iparties or

. . . . . . . . ]thefamily J o, I / 2~°atT~'r~mage [ 2 a I~ before o ~ t e r th~" ~ ~ . ld~ ~the ma~l~.~ 2. maintained ~ ~arriage (Th ~ / c le~ ....... ~0 0 21

,twas

'~'"'~"

~ I

!~ Conclusion"• The vehicle ! I will be included in the pool [ but the Court cannot order a I Irl transfer [ " 0222,0i222,02122,0112

~__..Z

-.._~_,;.__~,~_7,_o_,~

1

2 ~

~ ~ m~nB~

1. !he husband ~ 2 : . 92"I 2,011

1. not maintained or improved during the marriage

or the wife

I conclusion: r ~ vehicle ! [will be included in the pool [ I and the Court can order a [ I transfer | T 0221,01221,02121:0111 ," . ! Conclusion: The I . [ vehicle will be [ ~] excluded from the pool l 0121,021 !

Fig. 1. Directed graph for common pool determination of vechicles.

An arc in the graph may lead to a node on the same graph or alternatively it may lead to an independent graph. The capacity to leap to other graphs is required in order to implement not sure alternatives. For example, a user may be unsure as to whether or not the vehicle was a gift. An independent graph has been developed to resolve this question. The not sure option leads to this graph and returns with a conclusion from it.

4. The percentage split determination in Split-Up A rule-based approach for the percentage split deternlination is made difficult in that the knowledge necessary for such a determination cannot be obtained from any statutes in the z This program known as the BUILDER was developed using Microsoft's Visual Basic and runs on an IBM PC platform. 3 This program, called the RUNNER, uses a combination of SQL and procedural code to implement each of the set theoretic operators required for the inferencing and explanations.

A LEGAL EXPERT SYSTEM

271

Family Law Act (Cth.). Further, heuristics in the domain are complex and numerous, due to the discretionary nature of the Act. Whilst the Act specifies a number of factors that a judge must consider in determining a percentage split of the assets, it allows much latitude in the combining and weighting of these factors. Heuristics underlying discretionary acts are, in general, difficult to model with rules, essentially because a rule is required for every possible combination of factors and their values. A discussion of how to model discretionary acts can be found in [Zeleznikow et al. 1994]. Neural networks, in contrast to a rule based approach, determine and represent weights of factors sub-symbolically and thus are well suited to capturing the weighting of factors which predict a judge's performance. Factors specified by the statute as relevant for a percentage split determination can be selected as inputs into a neural network and the output can be the percentage of the assets awarded to the parties. A supervised network is preferred over an unsupervised network as the output, namely the percentage split reported in a judgement, is known in all cases. A pool of three hundred unreported cases serves for the extraction of case facts so that a training set may be assembled. Access to unreported cases was granted by the Family Court of Australia (Melbourne registry). In general, cases remain unreported if the judgement does not establish any significant legal precedents. Feed forward networks were trained using back propagation of errors. The software used was QuickProplI. QnickProplI implements an algorithm which selects optimal learning rates, momentum and bias terms during training. Other training algorithms require these values to be fixed before training, a requirement which is only successfully achieved by trim and error.

5. Explanation in Split-Up Because neural networks are incapable of providing an explanation of their reasoning process, we decided to use Toulmin argument structures to provide explanation in the Split-Up system. [Toulmin 1958] concluded that all arguments consist of four invariants: claim, data, warrant and backing. 4 A detailed descriptions of how Toulmin argument structures are used in Split-Up can be found in [Stranieri et al 1994b]. The assertion of an argument stands as the claim of the argument. The claim can be substantiated by data. The warrant acts as a justification for the claim given the data. The backing of an argument supports the validity of the warrant. Toulmin Argument structures have been used in the field of artificial intelligence and law to represent legal arguments by [Dick 1991] and by [Marshall 1989]. [Branting 1994] has proposed an extension of Toulmin warrants as a basis of a model of ratio decidendi. In fields other than law, [Johnson et al. 1993] discern five distinct types of expertise that correspond to five types of backing. [Bench-Capon et al. 1991] annotate clauses in their logic programs with literals that represents components of argument schema proposed by Toulmin in order to generate explanations. 4 Toulmin also proposed two additional invariants, a rebuttal and a modality. The development of a computational model which reasons with these invariants is the topic of current research.

272

JOHN ZELEZNIKOWET A L DATA

1. George has contributed much more than Martha 2. George has similar future needs as Martha 3. The marriage is of average wealth

.•

CLAIM George receives 65% of the property

WARRANT 1. Past contributions are rewarded 1. Section 79(4)(a-c) | 2. Section 7 9 ( 4 ) ( e } NoeIVNoeIFLC | 92-083 /

2. Future needs are taken into account if significant

3. The percentage is influenced by the contributions to a greater extant in wealthy marriages than is the case in less wealthy marriages

BACKING

Fig. 2. Percentagesplit argumentin Split-Up. Sixty four arguments were identified for the determination of an appropriate percentage split of the assets of a marriage. Many of these arguments produced claims which were in turn used as data for other arguments. All arguments contribute to a culminating argument (the percentage split argument), the claim of which presents a solution to the problem. This argument is represented in Figure 2. For many arguments, the claim in inferred from data values with the use of a neural network. The inputs into the network are the data items for the argument. The network's output represents the claim of the argument. For a number of arguments, the claim is inferred from data by the use of rules. The warrant and backing for an argument are applicable whether a neural network or rules have been employed. The percentage split module of Split-Up has been implemented using the object oriented knowledge based system development tool, KnowledgePro. The hypertext facilities built into KnowledgePro allow the warrant and backing based explanations to draw on statutes and past cases. Those arguments which are rule based make use of KnowledgePro's forward and backward chaining inferencing facilities. Cormectionist based arguments call on procedural functions that implement the feed forward component of neural networks. An explanation which is independent of the inferencing method used to produce the claim may be generated from the Toulmin structures as the following sample consultation illustrates. An explanation can be generated whether a rule set or a neural network has been used to produce the claim. Split-up:

Georgereceives 65% of the property. { Claim of Argument Percentage split}

User:

Why?{Claim is questioned}

A LEGAL EXPERT SYSTEM

273

Split-Up: Because George has contributed much more to the marriage than Martha. George has similar future needs as Martha. The marriage is of average wealth {Data of Argument Percentagesplit} User: So what? {Argument is questioned} Split-Up: Past contributions are rewarded and future needs are taken into account if significant. {Warrantof Argument Percentagespilt} User: Why? {Warrant is questioned} Split.Up: 1. Section 79(4)(a-c) of the Family Law Act explicitly directs that contributions are to be taken into account. 2. Section 79(4)(e) of the Family Law Act allows the Court discretion regarding the relevance of future needs. Noel and Noel (1981) FLC 92-083 advises that future needs should be considered if either party's needs are significant {Backing of Argument Percentagesplit}

6. Competing strategies for deriving Toulmin Argument Structures in Split-Up We use two different approaches in Split-Up to infer the claim of an argument: inferencing from data by the use of rules and by the use of a neural network. It is thus significant to examine the relative merits o f each strategy and determine which technique should be used for a given problem. To compare these strategies, we have implemented one argument using both rules and a neural network. This argument has as its claim, the capacity of an individual to engage in employment. A n individual has the capacity for full time employment, part time employment only or is not a b l e to be employed at all. Data for this argument includes the person's health, age and parenting responsibilities. The data becomes the input to the neural network and the claim is the output. The training set comprised 105 records which represent each possible combination of inputs. The outputs for 48 of these records were elicited from past case judgements. The remaining outputs were derived from hypotheticals which were determined in conjunction with the domain expert. The neural network trained to an acceptable error rate in less than 1300 epochs. Initially, eleven heuristic rules were elicited from the domain expert for this argument. In order to evaluate the agreement between the neural network output and the rule base output we used the initial 105 record training set as a test set. Results of comparison are illustrated in the following table. Table 1. Table comparing rules and neural networks for the capacity to work argument using original rules

Neural net Output

Full time Part time None

Rule Output Full time

(11 rules) Part time

None

30 0 0

0 30 5

0 7 33

The non diagonal entries indicate the number o f trials where the rule based output did not agree with the neural network output. Disagreement between the rules and neural network on these 12 trials was found to be due to an incomplete rule set. When two new

274

JOHNZELEZNIKOWET AL.

rules were added to the base, there was total agreement. The neural network thus acted to validate the heuristic rules. The rule set cannot reason with inputs which have more than one value for health, age or parenting. This occurs often and reflects the extent to which inputs are necessarily uncertain. For example, if health was considered excellent or good, and age was young or quite young, the neural network produced an output indicating the person has a capacity for full time work. Non fuzzy rules cannot reason with these inputs while the neural network always produces an output. The technique of extracting rules from data with the use of various induction algorithms has been used in a number of applications as an alternative to the elicitation of heuristics from experts. To compare the two approaches the training set data was passed through the ID3 induction algorithm. The ID3 algorithm produced 62 rules from the 105 record training set. This can be contracted to the eleven rules originally elicited from the domain expert, which were supplemented by the two additional rules. The large number of rules resulting from the use of the ID3 algorithm was not entirely unexpected. The assumption underlying the ID3 algorithm is that classes are defined by their attributes. As the ID3 algorithm continues to divide the input data set, the number of classes represented at each division is only reduced occasionally. This results in a very large decision tree and a large number of rules. It should be noted that the results of this paper are preliminary, and the research we have discussed here is continuing. A full paper exploring the competing strategies using bench-marks will be published at the conclusion of this project.

7. Using Split-Up to perform negotiation As we stated in the introduction to this paper, our original aim in this project was to support legal negotiation. We have been able to do so in two ways: - In [Meersman et al. 1994] we developed mathematical models to determine what are the litigant's beliefs and goals, how far apart their goals are, and what will be the cost to each party if they decided to litigate to achieve their goals. By commencing at the mid-point between the parties goals one could initiate a plausible negotiation process. - An alternative approach we have used is to feed into the Split-Up system both the plaintiff and defendants beliefs and goals and see what they would be likely to be awarded if the judge accepted their version of the facts. [Meersman et al. 1994] gives a detailed description of a case in which the wife wants 60% of the common pool and the husband requires 50% of the common pool. 5 Given each litigant's view of the facts the system awarded the wife 65% and the husband 58%. The Split-Up system allows users to play hypotheticals. If the wife were to receive custody of their two children; Split-Up suggests she would receive 60% of the common pool, even accepting the husbands view of the facts. Hence it may be appropriate for the husband to settle, especially if he believes that he is unlikely to win custody of the boys.

References Bench-Capon, T. J. M., Lowes, D. & McEnery, A. M. 1991. Argument-Based Explanation of Logic Programs. Knowledge Based Systems 4(3): 177-84. 5 They are also is disagreement as to the exact value of the common pool.

A LEGAL EXPERT SYSTEM

275

Branting, K. 1994. A Computational Model of Ratio Decidendi. Artificial Intelligence & Law 2: 1-31. Dick, J. P. 1991. Representation of Legal Text for Conceptual Retrieval. Proceedings of The Third International Conference on Artificial Intelligence and Law, 244-253. USA: ACM Press. Ingleby, R. 1993. Family Law and Society. Sydney: Butterworths. Johnson, P. E., Zualkeman, I. A. & Tukey, D. 1993. Types of Expertise: An Invariant of Problem Solving. International Journal of Man Machine Studies 39: 641-65. Kolodner, J. L. & Simpson, R. L. 1989. The MEDIATOR: Analysis of an Early Case Based Problem Solver. Cognitive Science 13: 507-549. McCarty, L. T. 1983. Intelligent Legal Information Systems: Problems and Prospects. Rutgers Computer and Technology Law Journal 9(2): 265-294. Marshall, C. C. 1989. Representing the Structure of Legal Argument. Proceedings of The Second International Conference on Artificial Intelligence and Law. USA: ACM press. Meldman, J. A. 1977. A Structural Model for Computer-Aided Legal Analysis. Rutgers Journal of Computers and Law 6:27-71. Meersman, R., Zeleznikow, J. & Hunter, D. 1994. A formal Model of Legal Negotiation Submitted to Fifth Intemational Conference on Artificial Intelligence and Law. Washington, USA. Ross, H. L. 1980. Settled Out of Court. Aldine. Smith, J. C. & Deedman, C. 1987. The Application of Expert Systems Technology to Case-Based Law. In the Proceedings of The First International Conference on Artificial Intelligence and Law 84-93. Boston: ACM Press. Stranieri, A. & Zeleznikow, J. 1992. SPLIT-UP Expert System to Determine Spousal Property Distribution on Litigation in the Family Court of Australia. In the Proceedings of The Fifth Australian Artificial Intelligence Conference 51-66. Hobart: World Scientific. Stranieri, A., Massey, P. & Zeleznikow, J. 1994. Inferencing with legal knowledge represented as diagrams. In Williams A. W. F (ed) Poster Proceedings of The Seventh Australian Joint Conference on Artificial Intelligence AI'94, p. 25-32. Stranieri, A., Gawler, M. & Zeleznikow, J. 1994b. Toulmin Structures as a Higher Level Abstraction for Hybrid Reasoning. In Proceedings of The Seventh Australian Artificial Intelligence Congress AI'94, 203-210. Armidale. World Scientific, Singapore. Sycara, K. 1990. Negotiation Planning: An AI Approach, European Journal of Operations Research 46: 216-234. Toulmin, S. 1958. The Uses of Argument. Cambridge: Cambridge University Press. Vossos, G., Zeleznikow, J., Moore, A. & Hunter, D. 1993. The Credit Act Advisory System (CAAS): Conversion from an Expert System Prototype to a C++ Commercial System. In the Proceedings of The Fourth International Conference on Artificial Intelligence and Law, 180-183. Amsterdam: ACM Press. Zeleznikow, J. & Hunter, D. 1994. Building Intelligent Legal Information Systems, Deventer: Kluwer Law and Taxation. Zeleznikow, J. Hunter, D. & Stranieri, A. 1994. Reasoning in Open Textured Domains: Benefits of Integrating Multiple Reasoning Stragies. In Bramer, M, A. & Macintosh, A, L. Research and Development in Expert Systems XI. Proceedings of Expert Systems 94, 187-198. Cambridge: SGES Publications.