Artificial intelligence (AI) has become a popular buzz phrase in the document capture market recently. At its root, AI means enabling machines to think like human brains. When applied to capture, AI typically incorporates elements like natural language processing (NLP) and semantic understanding.
While being able to think like a human may sound like an ideal way to process documents, it’s important to remember that one important way people learn is through their mistakes. And when applying document capture in mission-critical areas in markets like healthcare, financial services, and government, the tolerance for mistakes is very low. That is why Ephesoft has chosen to implement machine learning technologies, instead of AI, in its software.
Applied Machine Learning
Machine learning was originally defined as “the ability to learn without being programmed,” and that is exactly what we have done with our capture software. To identify classes of documents, as well as fields to be extracted, a user only has to present our software with examples of those classes and fields. Our software will take it from there, with some assistance from the user.
This ability to learn by example, vs. traditional capture applications that typically require setting up templates and writing rules based on regular expressions, greatly reduces set up times. And, while other vendors may advertise similar capabilities, our software’s ability to learn from just a few examples differentiates it from competitive products that often require hundreds of images be submitted as examples. By necessity, working with high volumes of examples makes their applications specialized and inflexible. (High volumes of examples are also typically required for AI-based technologies like NLP.)
Our software is enabled by its ability to analyze multiple characteristics on a document, such as the position of text on a page, the text fonts and sizes used, elements positioning related to other elements, and more. Overall, our patented technology can draw information from more than 20 elements in a document to enable it to do classification and extraction. Most competitive products can analyze a handful of elements at most.
Our software’s ability to learn from so few examples enables capture processes to potentially be set up by almost anyone in an organization to suit their personalized needs. In a bank, for example, a compliance officer and a marketing executive might be working with the same set of loan documents, but with completely different goals. Even two marketing professionals might have completely different routes they wish to take for capturing data for market analytics.
This is where the concept of “individual supervised” machine learning comes into play. Each individual has the ability to execute our classification and capture technology in the way they best see fit. To accommodate them, we avoid the concept of “clustering,” which basically means pre-grouping documents based on pre-defined characteristics. We enable each individual to set up and supervise groupings in their own fashion.
Users can then train our application to suit their needs as well, validating the results and making corrections when needed. Minimum confidence thresholds can be adjusted to ensure that no mistakes are getting through, and the software will learn from the corrections being made.
The bottom line is that supervised machine learning applications are designed to be trained without making errors, similar to the way a computer operates. This is the type of learning that best suits the mission-critical applications in which document capture typically plays.