ELIS Evidence Classifier Service
Description
The adjudicators and contractors spends too much time sifting through digital evidence documents for relevant information.
Detailed example
Tagged evidence. The system inputs an image (scanned document from Lockbox) and outputs either a specific label, such as "Border Crossing Card - Front," or no label if that document is not recognized as one of the classes.
AI / analytics pattern
Classical/Predictive Machine Learning: Models trained on data to make predictions or classifications based on identified patterns or relationships.
Automation level / stage
c) Deployed – The use case is being actively authorized or utilized to support the functions or mission of an agency.
Expected benefit
To enable end users to navigate directly to the page(s) containing evidence documents of interest instead of sifting through large PDF documents. Evidence tagging intends to accelerate case processing by identifying specific types of documents (e.g., I-589, passport photo spread, marriage certificate) and applying a metadata tag to that document object in ELIS. This way, when a user opens a case with potentially hundreds of pages of evidence documents, rather than scrolling through them one at a time to find a specific document of interest, they have clickable "bookmarks" in the UI generated from these tags that will jump directly to the corresponding page.
Controls / human review
ATO: Yes; PIA: https://www.dhs.gov/publication/dhsuscispia-056-uscis-electronic-immigration-system-uscis-elis
Data needed
The system consists of a single vision-based object recognition model, and many text-based binary classifiers. The text models were trained and evaluated on separate class-specific sets of production data sampled from evidence documents, and each data point is the linearized OCR text obtained from a single scanned page image and AWS Textract. These training and testing sets are then annotated by data scientists on our team.