On Design and Evaluation of High-Recall Retrieval Systems for Electronic Discovery

Abstract

High-recall retrieval is an information retrieval task model where the goal is to identify, for human consumption, all, or as many as practicable, documents relevant to a particular information need. This thesis investigates the ways in which one can evaluate high-recall retrieval systems and explores several design considerations that should be accounted for when designing such systems for electronic discovery.

The primary contribution of this work is a framework for conducting high-recall retrieval experimentation in a controlled and repeatable way. This framework builds upon lessons learned from similar tasks to facilitate the use of retrieval systems on collections that cannot be distributed due to the sensitivity or privacy of the material contained within. Accordingly, a Web API is used to distribute document collections, informations needs, and corresponding relevance assessments in a one-document-at-a-time manner. Validation is conducted through the successful deployment of this architecture in the 2015 TREC Total Recall track over the live Web and in controlled environments.

Using the runs submitted to the Total Recall track and other test collections, we explore the efficacy of a variety of new and existing effectiveness measures to high-recall retrieval tasks. We find that summarizing the trade-off between recall and the effort required to attain that recall is a non-trivial task and that several measures are sensitive to properties of the test collections themselves. We conclude that the gain curve, a de facto standard, and variants of the gain curve are the most robust to variations in test collection properties and the evaluation of high-recall systems.

This thesis also explores the effect that non-authoritative, surrogate assessors can have when training machine learning algorithms. Contrary to popular thought, we find that surrogate assessors appear to be inferior to authoritative assessors due to differences of opinion rather than innate inferiority in their ability to identify relevance. Furthermore, we show that several techniques for diversifying and liberalizing a surrogate assessor's conception of relevance can yield substantial improvement in the surrogate and, in some cases, rival the authority.

Finally, we present the results of a user study conducted to investigate the effect that three archetypal high-recall retrieval systems have on judging behaviour. Compared to using random and uncertainty sampling, selecting documents for training using relevance sampling significantly decreases the probability that a user will identify that document as relevant. On the other hand, no substantial difference between the test conditions is observed in the time taken to render such assessments.