Knowledge discovery for decision support in law

The Split Up project applies knowledge discovery techniques (KDD) to legal domains. Theories of jurisprudence underpin a classification scheme that is used to identify tasks suited to KDD. Theoretical perspectives also guide the selection of cases appropriate for a KDD exercise. Further, jurisprudence underpins strategies for dealing with contradictory data. Argumentation theory is instrumental for representing domain expertise so that the KDD process can be constrained. Specifically, a variant of the argumentation structure proposed by Toulmin is used to decompose tasks into independent sub-tasks in the data transformation phase. This enables a complex KDD exercise to be decomposed into numerous simpler exercises that each require less data and have fewer instances of missing values. The use of the structure also facilitates the development of information systems that integrate multiple reasoning mechanisms such as first order logic, neural networks or fuzzy inferences and provides a convenient structure for the generation of explanations. The viability of this approach was tested with the development of a system that predicts property split outcomes in cases litigated in the Family Court of Australia. The system has been evaluated using a mix of strategies that derive from a framework proposed by Reich. 1. RESEARCH OBJECTIVES AND QUESTIONS The Split Up project aims to apply knowledge discovery from database (KDD) processes to law. Specific research questions are: • Can theories of jurisprudence be applied to facilitate the selection of tasks in law that are suited to a KDD process? • Can theories of argumentation be beneficially applied to integrate expert knowledge into the KDD process? Knowledge discovery techniques have not been applied extensively in the legal domain despite potential benefits in the automated generation of legal knowledge from data. The absence of data in quantities collected in other fields, such as astronomy, in part accounts for this trend. However, for the most part, KDD has not been extensively performed with legal data because of a lack of clarity about how this can be done. Theories of jurisprudence have proven indispensable for the analysis and development of computational models of legal reasoning. For example, the rule positivism of (Hart 1961) underpins the application of logic programming in law exemplified by (Sergot et al. 1986). The identification of jurisprudence theories that are particularly applicable to improve KDD and how they can be applied is the first objective of the current research. In practice, a knowledge discovery from database process involves the incorporation of some domain expertise at each of the following KDD phases: data selection, preprocessing, transformation, mining, and evaluation. According to argumentation theorists, domain expertise can conveniently be represented as arguments for or against assertions. Therefore, we surmised that argumentation may provide a convenient framework for the representation of domain expertise so as to improve results from a KDD exercise. Stranieri and Zeleznikow 636 INFERENCE WARRANTS RELEVANCE WARRANTS neural network inference mechanism DATA CLAIM H has contributed X relative to the wife H has Y resources relative to the wife The marriage is of Z wealth Backing Why Data is relevant Section 79(4) Statute makes this relevant Section 75(2) Statute makes this relevant Lee Steere, Brown Precedent cases Backing Why inference is appropriate 103 unreported cases Network trained with appropriate examples Studies cited by Haykin Network is trained with a proven learning rule: BP Husband is likely to be awarded P percent of assets Figure 1. Toulmin Argument Structure for One of the Split up Arguments The Split Up project applies KDD to predict property division decisions made by judges of the Family Court of Australia following a divorce. We aim to develop strategies that are applicable to other areas of law and, more broadly, to other domains. In this paper, we report the following findings to date: • KDD is particularly suited to the discovery of knowledge in discretionary domains. Discretionary domains are defined by jurisprudence theories. • The discernment of tasks suited to KDD from those more appropriately suited to other methods relies heavily on the jurisprudential concept of open texture. • The argumentation theory proposed by Toulmin (1958) is used for the representation of domain knowledge in the data transformation phase. Further, the argument structure we use also facilitates the development of hybrid systems and the generation of explanations In the next section we elaborate on the theoretical foundations that underpin these points. 2. THEORETICAL FOUNDATIONS The concept of open texture is prevalent in law. This concept was introduced by Waismann (1951) to assert that empirical concepts are necessarily indeterminate. To use his example, we may define gold as that substance which has spectral emission lines X and is colored deep yellow. However, because we cannot rule out the possibility that a substance with the same spectral emission as gold but without the color of gold will confront us in the future, we are compelled to admit that the concept we have for gold is open textured. A definition for open textured terms cannot be advanced with absolute certainty unless terms are defined axiomatically, as they are, for example, in mathematics. The concept of open texture is significant in the legal domain because new uses for terms and new situations constantly arise in legal cases. Prakken (1997) discerns three sources of open texture: reasoning, which involves defeasible rules; vague terms; or classification ambiguities. We add judicial discretion as conceptualized by Christie (1986) and Bayles (1990) to this list and argue that judicial discretion is a form of open texture distinct from the three listed above. According to this view, decision makers have some flexibility in weighing relevant factors when exercising discretion although articulating an assignment of weights is typically difficult. The KDD process is particularly useful for discovering the weights of relevant factors from a database of decided cases. Domain expertise in family law is represented in the Split Up system as arguments. This enables an informed data transformation phase and also constrains the data mining. For the philosopher Toulmin (1958), practical reasoning, as distinct from analytical reasoning, involves the construction of an argument. Arguments, regardless of the domain, have a structure that consists of six basic invariants: claim, data, modality, rebuttal, warrant, and backing. Every argument makes an assertion based on some data. The assertion of an argument stands as the claim of that argument. A warrant justifies why the claim follows from the data. The backing supports the warrant and in a legal argument is typically a reference to a statute or a precedent case. The rebuttal component specifies an exception or condition that obviates the claim. The Toulmin argument structure has been used by a number of researchers in various fields to model reasoning. However, a survey by Stranieri and Zeleznikow (1999) illustrates that the majority of researchers vary the structure to suit their particular use. The variation that we used aimed to facilitate KDD. The structure is illustrated in Figure 1. In the next section, we shall discuss how open texture and argumentation have been applied to legal reasoning in the Split Up project. Decision Support in Law 637 Percentage split task given contributions and needs Narrow unbounded Wide unbounded Narrow bounded Wide bounded