Understand Assessment in Relevance

Note: Advanced eDiscovery requires an Office 365 E5 subscription for your organization. If you don't have that plan and want to try Advanced eDiscovery, you can sign up for a trial of Office 365 Enterprise E5.

Advanced eDiscovery enables early assessment, for example, for the defined issues and the data imported for a case. Advanced eDiscovery enables the expert to make decisions pertaining to an adopted approach and to apply them to the document review project.

Understanding assessment

In Assessment, the expert reviews a random set of at least 500 files, which are used to determine the richness of the issues and to produce statistics that reflect the training results. Assessment is successful when enough relevant files are found to reach a statistical level that will help Advanced eDiscovery Relevance to provide accurate statistics and to effectively determine the stabilization point in the training process.

The higher the number of relevant files in the assessment set, the more accurate the statistics and the effectiveness of the stability algorithm. The number of relevant files within the assessment files depends on the richness of the issue. Richness is the estimated percent of relevant files in the set relevant to an issue. Issues with higher richness will reach a higher number of relevant files more quickly than issues with lower richness. Issues with extremely low richness (for example, 2% or less) will require a very large assessment set to reach a significant number of Relevant files.

The statistics, which are presented in the Track and Decide tabs during training and after Batch calculation, include estimations of recall for different review sets. In statistics, estimations that are based on a sample set (in this case, the assessment files) include the margin of error and the confidence level of that error margin. For example, estimated recall of 80% might have a margin of error of plus or minus 5% with a confidence level of 95%. This means that the estimated recall is actually 75%-85% and this estimation has 95% confidence. The larger the assessment set, the margin of error becomes smaller and the statistics are more accurate.

After the expert reviews an initial assessment set of 500 files, Relevance is able to determine the current margin of error of the recall values. Relevance will also set a default margin of error that it recommends to reach to optimize the assessment set. Following are some examples:

  • If the assessment set already yielded a margin of error of plus or minus 10%, Relevance will recommend to move on to training (no additional assessment review is needed).

  • If the assessment set yielded a margin of error of plus or minus 13%, Relevance might recommend the review of another set of assessment files to reach a smaller margin.

  • If richness is extremely low, Relevance might recommend stopping assessment even though the margin of error is large (making statistics impractical), because the assessment set needed to reach a useful margin of error is too large.

Each issue has its own richness, current margin of error, and as a result, estimated number of additional assessment files. The next assessment set is created according to the maximum number of files (up to 1,000 in a single set).

You can accept the Relevance recommendations or adjust the current margin of error according to your needs. The default current margin of error is determined for recall at equal or above 75%.

Notes: The Assessment stage can be bypassed, in the Relevance > Track tab in the expanded view for an issue, by clearing the Assessment check box per issue and then for “all issues”. However, as a result, there will be no statistics for this issue.

Clearing the Assessment check box can only be done before assessment is performed. Where multiple issues exist in a case, assessment is bypassed only if the check box is cleared for each issue

See Also

Office 365 Advanced eDiscovery

Tagging and Assessment

Tagging and Relevance training

Tracking Relevance analysis

Deciding based on the results

Testing Relevance analysis

Share Facebook Facebook Twitter Twitter Email Email

Was this information helpful?

Great! Any other feedback?

How can we improve it?

Thank you for your feedback!

×