Expediting a routine case turned urgent
In this matter, Phillips Lytle represented an employer in a discrimination case involving a universe of over three million documents. Initially, a traditional workflow of linear review and search terms was anticipated given the straightforward nature of the project. Culling with search terms reduced the universe from millions down to a more reasonable half-million and the team then began the manual review. However, just a few thousand documents into the review, the team was hit with an expedited deadline that moved the timetable up from months to just weeks. Relying on a linear review process was no longer an option.
Sampling for intelligent and defensible project planning
The Phillips Lytle team engaged OpenText™ Discovery to leverage Axcelerate for the review with support from the OpenText Professional Services team lead by Walker Hartz and his client services group. The unanticipated change in circumstance necessitated an innovative approach, combining statistical sampling, search terms, AI and validation testing to complete the review in time. Axcelerate’s continuous machine learning algorithm, predictive coding, was crucial to the process. It learns from human reviewer decisions, identifies the common characteristics of relevant documents and then finds similar content automatically, categorizing it and prioritizing the most likely relevant documents for human review. First, the team needed to understand the scope to craft an appropriate attack plan. Using Axcelerate’s integrated random sampling tools, the team was able to identify a reliable responsive rate of around 1.5 percent. In other words, the team knew they were looking for approximately 10,000 documents.
Repurposing existing work product and review decisions
To kickstart the machine learning process, the team repurposed roughly 5,000 decisions from the manual review. Additional documents identified through pattern-recognition and phrase analysis were added to the training cycle. This data was more than enough to generate reliable training for Axcelerate’s predictive coding engine. The initial training round ranked the potential relevance of the entire corpus of data on a scale of 1-100, and the team began carving up tranches for further investigation. Axcelerate’s unique approach to predictive coding enables user control when the system retrains and re-ranks documents. This functionality provided the team with the freedom to experiment and identify the sweet spot for highly relevant batches.
Applying predictive coding for expedited production
The team designed a unique workflow that hybridized two of the most widely used and judicially approved TAR approaches. Instead of coding a single training iteration, such as a seed set, entirely in the dark, the team leveraged TAR 2.0 to continually learn and suggest potentially relevant documents throughout the process, generating a far more accurate seed set than is possible under traditional TAR 1.0 workflows. After several rounds of targeted training, informed by statistical samples at each stage, the team used AI to categorize the remaining documents and bucket them into discrete confidence levels (e.g., 60 percent confident, 70 percent confident).