Models for Predicting Legal Outcomes

INTRODUCTION Using a database of cases represented as sets of features and outcomes, computer programs can predict the results of new problems. They do so usually by means of case-based reasoning (CBR) models or ML algorithms, sometimes combining the two. This chapter will explain and illustrate both. The prediction techniques make use of different types of features represented in prior cases, ranging from names of judges deciding and law firms litigating cases of different types, to attitudinal information about judges, to historical trends in decisions, to stereotypical fact patterns that strengthen a claim or defense, that is, legal factors. Such features differ in the extent to which they capture information about the merits of a case. Judges’ names, law firm names, and type of case, for instance, patent litigation or product liability, capture no information about the merits of a particular legal dispute. In contrast, legal factors, as we have seen in Chapter 3, capture quite a lot of information about the merits. Such features also differ in terms of the ease with which they can be extracted automatically from the texts of cases. Judicial and firm names are easy to extract; legal factors can be extracted but it is considerably harder to do so. This book focuses on features that tend to capture some information about the merits of a case and that feasibly can be extracted automatically from case texts. This chapter explores alternative methods of using such features to predict case outcomes. Machine learning techniques use feature frequency information statistically to “learn” the correspondence between case features and target outcomes. Case-based techniques are focused more on case comparison and explaining predictions. They make predictions based on the strengths of competing legal arguments. The techniques vary in the ways in which they can explain their predictions and in the extent to which their feature representations are compatible with cognitive computing. Questions this chapter addresses include: How can computer programs learn to predict case outcomes? What is ML? What are supervised ML and decision trees? What are the advantages of random forests of decision trees? How does a CBR approach to prediction differ from ML? Will legal practitioners accept legal predictions without explanations? How can a program pose and test a prediction hypothesis? How can such prediction programs be evaluated empirically?