Putting names to the numbers
Prior to the new solution, an Arolsen Archives volunteer needed roughly 15 minutes to extract and upload each document. With the new AI-based approach, it takes less than 20 seconds. Since the implementation of the solution, more than 160,000 names have been indexed, more than 18,000 documents have been extracted, and more than 63,000 documents have been clustered, meaning that similar documents have been grouped together for easier and more accurate readings. On average, it takes less than one second per document to cluster, and it will only get faster over time as the AI continues to learn.
Despite AI’s being able to do roughly 95% of the work, there still needs to be a human element in the validation of documents. By bringing humans and machines together, a single volunteer (instead of 10) can now get through roughly 41 documents each hour. Freeing up time for the other volunteers to do the same, Arolsen has seen a 40-fold increase in productivity.
As for the AI’s confidence, it’s steadily rising. For instance, within the field of “mother’s last name,” the AI has gradually improved its confidence by 10% thanks to inputs from volunteers. When it comes to “religion,” the AI is operating at 99% confidence.