The Structure of Predictive Coding: A Guide for the Perplexed

From this perspective, the core attribute of PC is the de-linking of volume and cost, while maintaining or improving quality. But there are other valuable perspectives on PC. At first glance, the range of available PC systems and solutions is bewildering. A closer look reveals that the landscape is governed by a small number of structural choices—choices posed by the problems that predictive coding confronts. The goal of this paper is to isolate these structural choices and to show how particular responses to them provide a characterization of both the predictive coding systems themselves and also their range of applications. One justification for this goal is simple: characterizing the structural choices of PC solutions yields a clearer delineation of the range of systems available in principle and a clearer understanding of the different advantages and disadvantages of the particular systems available in the current e-Discovery marketplace. A core reason to focus on the range and compatibility of structural choices rather than particular technologies is that some of the relevant technology components play different structural roles in different structural configuations. A note on terminology: the terms ‘predictive coding’, ‘technology assisted review’ (TAR), and sometimes ‘computer assisted review’ (CAR) are at times conflated. In what follows, we use the term ‘predictive coding’ in a specific sense to denote review systems and protocols in which one or more documents are coded or labeled with review-sensitive information (‘responsive’ (R), ‘non-responsive’ (NR), etc.) without direct document-particular human supervision. (Of course, valuable systems of predictive coding will apply non-individual document coding to many more documents than just one. But we can draw a principled line at a single document.) In contrast, in human linear review—non-predictive coding—every label on every document is supplied through direct human judgment. In this terminology, then, ‘predictive coding’ stands in direct contrast to human linear review. If we take ‘technology assisted review’ and ‘computer assisted review’ in their natural interpretations, these phrases apply to any form of review assisted or aided by technology or computation—including many systems that do not apply any form of predictive coding. (Anyone using MD5 hashes for review deduplication is using TAR/CAR. But they’re not necessarily using PC.)