Review Center Guide 9
The AI classifier uses the extracted text of documents to make its predictions. Even if other fields are
returned in the saved search, it will not affect the results.
If you choose a prioritized review queue, we recommend coding at least two non-empty documents in your
data source before preparing or starting the queue: one with the positive choice on your review field, and
one with the negative choice. This gives the AI classifier the information it needs to start making its
predictions. The more documents are coded, the more accurate the classifier’s predictions will be.
If you do not have any coding completed, you can start the prioritized review queue without any coding. The
classifier model won't build until at least 50 documents have been coded, with at least one coded positive
and one coded negative. After you reach 50 coded documents, your ranks will update upon the next auto-
refresh or manual refresh. If you need it to build sooner, you can manually trigger a queue refresh at any
point after at least one document has been coded positive and one has been coded negative.
2.2.2.1 Including random documents in the queue
When you set up a prioritized review queue, you have the option to serve up randomly chosen documents
alongside documents that are predicted relevant. This gives the AI classifier a broader variety of coding
decisions to learn from, which improves its predictions in the early stages of a review. Having reviewers
code a selection of random documents helps the classifier identify a wider range of relevant topics and
prevents it from focusing on a limited subject area.
Under the queue setting Include Random Items, you can choose to include random documents as up to
20% of the total documents served to reviewers. You can change this setting at any time. We recommend
including a high percent of random items during the early stages of review.
2.2.2.2 Using Coverage Mode
When Coverage Mode is turned on for a prioritized review queue, the queue switches away from serving up
the highest-ranking documents. Instead, it serves up documents that are better for training the model.
These are documents with scores near 50, which usually have different content and topics from documents
that the model has previously seen. Labeling these helps the model learn from a wider variety of documents
and become more confident quickly.
When in Coverage Mode, the AI classifier sorts all documents by their scores’ distance from 50, but limits
and spreads out the number of exactly 50-ranked documents. This intermixing diversifies the group of
documents and lowers the chance of duplicates. The classifier then serves up these sorted documents to
reviewers until the next refresh. After each refresh in Coverage Mode, it re-sorts the documents. Coverage
Mode also overrides the Include Random Items setting.
You can turn the Coverage Mode setting on or off at any time during a review. For instructions, see Turning
Coverage Mode on and off on page19.
Note: Whenever you turn Coverage Mode on or off, manually refresh the queue. This updates the
document sorting for reviewers. For more information, see Turning Coverage Mode on and off on
page19.
2.3 How document assignment works
By default, five documents are checked out to each active reviewer at a time. As the reviewer saves their
progress on those documents, more are checked out as needed.
For example, documents 1 through 5 are assigned to the first reviewer who starts review. If a second
reviewer logs in immediately after, documents 6 through 10 are assigned to the second reviewer. As the first
reviewer completes their work, documents 11 through 15 are assigned to them, and so on.