Many solutions in the market promise automatic data extraction, yet they are based on templates or rules. Those rules and templates require setup and constant adjustments. Changing documents cause errors. Our deep learning algorithms infer the underlying structure of documents and thus do not require setup and but can automatically adjust to new documents.

Conventional solutions for automatic document processing work by means of applying fixed rules or templates previously defined by the user, with little to no ability of adaptation when it comes to new types of documents. In contrast, our solution fuses the visual features of the document together with its textual information in order to identify relevant regions and extract the different units of information contained within them.

Only the approach of understanding the underlying structure of documents allows for knowledge transfer across documents and corresponding accurcacy rates in information extraction. By the nature of their design, the above mentioned conventional solutions cannot reach a point where operators are mostly freed up from their work as they have to adapt the templates or rules with every change in the documents.