Information Extraction From Digital Receipts and Bank Transactions Using Machine Learning

This paper presents a machine learning solution so as to automate extraction of key information from submitted bank transaction portable document file (pdfs) such as (National Electronic Funds Transfer) NEFT acknowledgements and digital transaction receipts. The paper proposes a pipeline to preprocess the data sample to improve Tesseract's performance for text extraction. Key information extraction from digital receipts is implemented using two approaches namely the CUTIE model, a CNN model exploiting both spatial and semantic information and an objection detection model. Information extraction from bank transaction acknowledgements is achieved by pandas dataframe manipulation coupled with substring search and Fuzzy string matching. The cross-entropy loss function for the CUTIE model converges at 0.4. The F1 scores of the object detection model was 0.92 for amount class and 0.87 for other classes respectively.