Abstract: The disclosure describes a method including: obtaining a document image containing fields spatially arranged along multiple dimensions; determining first locations within the image, each of the first locations identifying a spatial position in the image of a respective one of the fields; extracting tokens from the fields; generating using a first trained model, based at least on the image, the first locations, and the tokens, first predictions, each first prediction indicating a classification of a respective one of the tokens; generating, using a second trained model, based at least on the image, second predictions identifying second locations, each of the second locations identifying a spatial position in the image of a respective entry of multiple entries, each of the entries including one or more rows of tokens; and linking, according to the first locations and the second locations, each of the tokens to one of the entries.

Abstract: Systems and methods to mask Personally Identifiable Information (“PII”) without masking proximate context are provided. The system may create a raw data set using at least one data file including PII through an ingestion pipeline. The system may identify and categorize the type of PII to be masked and record the type of PII in a data field. The system may selectively mask the PII by analyzing at least one location to determine a context proximate to the PII and masking only the PII. The system may output a masked data set including the at least one data file with the PII selectively masked and the data field containing a record of the type of PII selectively masked.