Information extraction from financial documents

Katsiaryna Stalpouskaya; Pavel Schkadzko; Katharina Robb; Karl-Heinz Krachenfels

Information extraction from financial documents

EuroSciCon Conference & Expo on Robotics, Automation & Data Analytics
April 08-09, 2019 | Paris, France

Katsiaryna Stalpouskaya, Pavel Schkadzko, Katharina Robb and Karl-Heinz Krachenfels

Gini GmbH, Germany

Posters & Accepted Abstracts: Am J Compt Sci Inform Technol

Abstract

In Fintech world a lot of transactions and operations still involve paper. For instance, for accounting or even private payments invoices are widely used. However, for further processing these documents need to be digitalised: for accounting all the relevant information from an invoice has to be put into accounting software by company employees; end users - in order to pay an invoice - have to type information manually in online banking app in order to transfer money. At Gini, we eliminate or minimise manual work by providing solutions for smart document processing by the means of AI. Our system takes a document in PDF format or as an image (taken by a hand-held camera or scanned) as input and outputs information necessary for an operation at hand, e.g., in order to complete a payment we extract IBAN number, the recipient of a payment, the purpose and the amount. For accounting purposes we identify dates, amounts (including amounts with and without tax), document number, names and addresses. Document processing pipeline can be split into three major steps. The first step is image processing. It is crucial especially for images taken with a mobile phone camera. In this step the document gets cropped, rectified and de-blurred. Thereafter the document is run through an OCR engine which provides actual text along with its location (words coordinates) in the document. Finally, we apply a recurrent neural network to distil relevant pieces of information, i.e. which strings produced by OCR represent necessary information (e.g., IBAN number, amount that has to be paid, etc.). In this talk, I will share the learning’s we gained in extracting information from document images with neural networks.