This is the final article in our five-part series highlighting the advancements in data extraction solutions over the past two years. Through this series, we aim to demonstrate the progress made in recent months and showcase the capabilities that were previously unavailable. The most notable improvement lies in the ability of current solutions to learn autonomously.
If you still need to read the introduction to this series, you can find it here: Mini-series: The real power of new data extraction solutions. The previous article discussed how modern solutions accurately handle multiple items within a list: Use Case 4 - Grouping Duplicates on Receipts.
Putting things in context
This article focuses on how recent data extraction solutions can effectively utilise all visual data present in documents to extract the information accurately. They use colours, font, font size, and layout to recognize the document and spot the information it needs to extract. Old solutions are extracting everything and trying to find the information in the extracted mess. Recent solutions recognize the document (global vision), search for the information to be extracted (selective attention) and extract - only - what they are looking for. Exactly like humans, they search for the data they are looking for, but much faster and more accurately.
Let's see, for instance, how they would recognize a company and all its associated information solely through its logo. On an invoice, just by the logo they identify the company that emitted the document. Beyond the technical feat, it has a significant advantage: it eliminates the need to extract individual data such as the company name and address. This reduces the probability of errors and the need for manual correction.
Marketing would agree: It’s all about the logo!
Let's see this in action. Consider a scenario where you need to automate payment to your providers in an account payable situation. The first step is to extract the name and contact information of the provider who issued the invoice. After all, this is where your company's money will be transferred. You'd better get it right!
There are two methods for extracting the provider's name on an invoice. The logo, which is usually prominently displayed on the document, is the first option. The second option is to extract the name and contact information from the bottom of the page, typically in tiny characters. OCR (Optical Character Recognition) relies on this latter information. However, due to the small size of the characters, achieving high accuracy requires intensive calculations or extensive post-processing checks.
A picture is worth thousands of words.
If a human were asked to identify the provider on an invoice, they would likely start by checking the logo and use the small character information as a check. Our recent data extraction models work in the same way. These models utilize all available visual information on the document, including layout and images, just as a human would. Consequently, the accuracy of extracted data is significantly improved, and the effort required to deploy and maintain these solutions in production is considerably reduced.
Since OCRs are prone to character-level errors, these solutions from the past had to reconcile several data to deduce the provider with a satisfying level of accuracy. Identifying a company with a high level of accuracy when the name, address, and VAT number contain a few typos takes work, even for a human.
By considering the layout of the document as well as the logo, the latest data extraction solutions get rid of these challenges. They output this critical information with high confidence in a single step.
A simpler workflow leads to fewer errors and less manual verification. Lower development costs, shorter deployment time, and lower maintenance costs mean a better return on investment for our customers.
Open your eyes!
If you would like to witness these solutions in action, please don't hesitate to contact us. We would be delighted to demonstrate the efficiency of these advanced data extraction solutions.
