AIDOX – Document Verification System

The existing document understanding systems use machine learning methods, natural language understanding and text analysis, to validate structured trade contracts for language and economic term correctness. The system is now being expanded to allow general document understanding across a wide variety of financial documents, beginning with a focus on customer provided reference material. The proposed solution will extract the key data elements from the documents and validate this data against the internal source of record. This provides several new challenges in document classification, visual document understanding and entity extraction. For structured documents a template-based solution is to be employed to extract key elements. For semi/un-structured documents, a multi modal framework is used incorporating text, layout and image information to extract key elements. Current state of the art solutions (BERT, LayoutLM) still fall short of human abilities, however by properly constraining our problem space we strive to achieve better results.

Faculty Supervisor:

Frank Rudzicz

Student:

Partner:

Scotiabank

Discipline:

Computer science

Sector:

Finance and Insurance

University:

University of Toronto

Program: