Inducing Predictive Models for Decision Support in Administrative Adjudication

Abstract

Administrative adjudications are the most common form of legal decisions in many countries, so improving the efficiency, accuracy, and consistency of administrative processes could significantly benefit agencies and citizens alike. We explore the hypothesis that predictive models induced from previous administrative decisions can improve subsequent decision-making processes. This paper describes three datasets for exploring this hypothesis: motion-rulings, Board of Veterans Appeals (BVA) decisions; and World Intellectual Property Organization (WIPO) domain name dispute decisions. Three different approaches for prediction in these domains were tested: maximum entropy over token n-grams; SVM over token n-grams; and a Hierarchical Attention Network (HAN) applied to the full text. Each approach was capable of predicting outcomes, with the simpler WIPO cases appearing to be much more predictable than BVA or motion-ruling cases. We explore several approaches to using predictive models to identify salient phrases in the predictive texts (i.e., motion or contentions and factual background) and propose a design for incorporating this information into a decision-support tool.