What does “handwriting recognition” really mean and what is the best way to practically apply it?
Handwriting recognition in Intelligent Document Processing (IDP) has undergone some exciting developments. This is primarily due to the ability to amass large amounts of sample data along with deep learning neural networks as well as available computing power to crunch and make meaning out of all of it.
If your organization has tried handwriting recognition before (or maybe its “cousin” intelligent character recognition or ICR) and not had much success, you definitely should try it again if you have even 10% of your data in handwritten form. The reason is that the range of applications has expanded while the prerequisites for configuring systems has been significantly reduced.
For example, only a few years ago, the best applications for handwriting recognition where those that could be easily constrained. This means for data that has a limited number of values or formats so that recognizers could make decisions using this additional context. The reason context was necessary is that, unlike OCR, which works on easy-to-learn fonts, handwriting recognizers had to decipher millions of different human fonts rendering the range of answers almost meaningless.
Leveraging Deep Learning
Now with the ability to use a lot more sophisticated deep learning algorithms on cheap – yet powerful – computing platforms, we’re not far off from the ability to translate, with a high degree of accuracy, pages of handwriting without any pre-configuration. This sounds exciting, but how is it practically implemented within intelligent document processing? There aren’t that many applications that require access to literal transcriptions. Rather, organizations need to operate on specific structured data so that business applications can use it. Turning unstructured data into structured formats of specific information is the realm of IDP.
The value of an IDP solution is primarily judged by one factor: the ability to convert as much unstructured data into structured information at a high degree of accuracy. For example, does plugging into Google cloud vision solve your problem? If you need access to structured information, the answer is almost 100% “no.”
Handwriting Recognition and Straight Through Processing
Is handwriting recognition, even highly accurate handwriting recognition, going to deliver significant amounts of straight through processing? That depends. It depends on whether you’re pulling-out data from structured forms, performing unstructured word-spotting, or trying to turn transcripts of handwritten text into some sort of meaning using natural language concepts.
The art of handwriting recognition within IDP, it turns out, is the same as the art of using text within IDP: a lot of consideration is required on what needs to be done to convert a transcript of handwriting or text into reliable, structured data.
While Parascript provides handwriting recognition that uses deep learning to provide accuracy levels better than anyone (even Google), we employ a lot of our expertise in the art of knowing where to look, how to convert that data into useful and usable structured information, and even how to create software that knows when it is correct and when to ask for assistance.
There is as much art to it as there is science. And using both, we deliver.