We just held a webinar in which Alan Pelz-Sharpe from Deep Analysis and I discussed the capabilities of machine learning-based IDP software and the use cases that can now be addressed. Alan provided a very useful construct to frame his discussion regarding a number of aspects of modern IDP software that enable the automation of processes, which were once untouchable. These include the ability to dramatically minimize the need to require human intervention, understanding complex information via context, the ability to learn and improve and work in suboptimal conditions. I’ll cover each at more detail in separate articles, starting with the last.
The phrase “working in suboptimal conditions” sounds like it should be associated with a Mars rover mission or OSHA litigation, not IDP software. In the case of document automation, “suboptimal conditions” is better defined as either poor data quality or highly-complex data.
Suboptimal Conditions: Poor Data Quality
On the image quality front, traditional IDP software employed a range of “image perfection” techniques performing tasks like detecting if an image was upside down, skewed in one direction, had excess noise or poor contrast. Advanced techniques included capturing color images to remove the background structure of a form (often referred to as drop-out forms) or to remove defects such as coffee stains or other image-based noise that could adversely affect OCR.
Today, there are still “poor working conditions” when it comes to image quality including images taken from a smart phone or documents transmitted by fax machine. Both of these use cases often result in images that have a significant amount of distortion that can render automation useless. For instance, fax machines can stretch or rescale images and introduce a lot of noise. Additionally, images that retain underlying preprinted form structure interfere with intolerant OCR engines.
For instance, faxed health claims is a well-known problem that continues to plague the insurance industry. The cost of processing even 10% of claims that are faxed often outweighs the cost of claims processing for the other 90% and renders auto-adjudication workflows useless without the aid of manual data entry.
IDP Improving Image Quality
With new deep learning image perfection, it is possible to train systems on a number of problems to correct for them. Systems can now take a poor quality faxed claim form and transform it into the equivalent of a high-quality, standard drop-out claim image, scrubbing the image of all the underlying form structure, resizing and rescaling the image, and leaving only the pristine field values for converting into machine readable text. Just take a look at this example.
Suboptimal Conditions: Highly-complex Data
As for working with very complex data, IDP software has evolved by leaps and bounds over solutions from even a few years ago. For instance, any process involving handwriting used to require a significant amount of effort to achieve even 50% automation. The most approachable use cases involved a very constrained scope where the system could be configured to anticipate what type of data was involved such as check amount recognition, address location and recognition for post, and structured forms recognition on highly-standardized fields such as dates, amounts and names.
Any need to transcribe multiple lines of handwriting was out of the question such as what is often the case for medical records: surveys where comments are allowed, and correspondence associated with claims litigation. For these broader data transcription needs, there was just no way to supply the system with advanced knowledge of the nature of the information.
IDP and Deep Learning
Enter deep learning neural networks that can deal with a lot more information and variance. This is the perfect answer to highly-variable handwriting because there is a distinct handwriting “font” for every person. Due to training deep learning algorithms on millions of examples, it is now possible to transcribe lines and even pages of handwriting fairly reliably.
While not as precise as OCR is on machine-printed text, it is getting there. The result is that these previously untouchable use cases where handwritten information must first be transcribed and then analyzed are now within the realm of possibility. Health records can be searched for individual conditions and recommended care. Insurance claims can be analyzed to understand the criticality of the situation.
Even though document capture has been around for decades——just like the venerable automobile——new technologies such as autonomous driving can seriously change the entire ballgame.