TAR 2.0

Technology Assisted Review (TAR) is a method in eDisclosure which expedites the document review process using machine learning.

There are two types of TAR:

1. TAR 1.0 is the traditional TAR which involves training the software with a set of documents, known as a “seed set”, which reviewers code as relevant or not relevant. The process requires constant iterations to be run by project managers to ensure the model understands any changes to decisions and what might be relevant next.

2. TAR 2.0 (better known as Continuous Active Learning - CAL) is a more advanced iteration and has overcome many limitations of TAR 1.0. Unlike its predecessor, TAR 2.0 does not include the hassle of needing project managers to rerun iterations and seamlessly learns from reviewer feedback much faster. This means the system can dynamically adapt to evolving changes and it does not matter when the last iteration was run. However, it is still advisable that project managers are utilised to run QC checks and ensure the model is performing as intended.

Therefore, CAL is a form of TAR that identifies and prioritises documents most likely to be relevant in a case. Unlike traditional TAR methods, CAL uses a continuous learning model, learning from reviewers’ coding decisions and continually improving its accuracy in identifying relevant documents. This significantly reduces the volume of documents that require manual review.

Imagine a legal case where there are a million documents to review for relevance. Reviewing these manually would be time-consuming and costly. This is where CAL comes into play. Initially, a small set of documents is reviewed and coded for relevance for the CAL model to initiate its build process, so that it can begin to understand and predict the next document which it considers to be relevant.

As the manual review exercise proceeds, the review queue is continuously reshuffled and reprioritised throughout the process (e.g. at a maximum frequency of every 20 minutes) as the system learns more about what is relevant and adjusts its predictions accordingly. If an initially low-ranked document is marked as relevant by a reviewer, the system adapts and serves up similar documents in the queue for review. This way, CAL ensures that even if some relevant documents are initially missed, they are likely to be caught in subsequent iterations.

The documents reviewed first are the ones which are most likely to be relevant. The documents remaining in the review pool are those most likely to be irrelevant and can be reviewed by a more junior reviewer or excluded from the review altogether. In the latter scenario, it is standard practice to run an Elusion Test to validate the accuracy of the review and to estimate how many low-ranked documents are actually highly relevant documents that you would leave behind if you stopped the project at that point.

This continuous learning and adaptation make CAL a powerful tool in eDisclosure, saving significant time and resources while increasing accuracy.