How intelligent capture works: Part 2 — Learning By adapting to variation and learning over time, intelligent capture is the starting point for generating clean and accurate data to fuel intelligent automation.

Missed Part 1? Find it here.

Traditional document capture — mainly associated with scanners, multi-function printers, and optical character recognition (OCR) — has been around for decades. Generally, these tools facilitate the transition of physical documents to fixed images from which OCR software then reads and extracts text.

(Fun fact: the concept of digitally converting printed text dates all the way back to 1914, leveraging telegraphy and reading devices for the visually impaired. More on that here.)  

The problem with traditional capture and OCR is that they need very specific conditions to be effective at reading text. For one, you need to start with a clean document.

Good luck with that.

Then, these tools can only read machine-printed text, maybe even only in specific fonts. To find the right text to extract and appropriately index a document, you’ll have to create a specific template for each document type or identify zones on the page where the OCR should pick up the text.

In other words, variation need not apply.

Capture tools are evolving and becoming more intelligent

Enter intelligent capture – also called cognitive capture. Boosted by technology advancements like intelligent algorithms and machine learning, cognitive capture tools classify documents and read text (even handwritten text) on documents with more variation. Cognitive capture tools also learn from the documents they read and process to improve the accuracy of classification and extraction over time.

Based on the data they pull from documents, intelligent capture tools may even trigger downstream tasks. In the case of accounts payable for example, they can automatically flag a staff member to create a new vendor record or post final transaction data to an ERP.

Put simply, intelligent capture is the starting point for generating clean and accurate data to fuel any intelligent automation solution or process.

It’s these advanced capabilities and built-in intelligence that set Brainware by Hyland apart and make it such an effective tool. It’s powerful because it leverages a neural network of intelligent algorithms that:

  1. Improve image quality
  2. Learn over time
  3. Understand imperfection

In my previous post, we explored how Brainware cleans up the noise to optimize document image quality prior to the OCR process. Now, let’s take a closer look at how it learns.

The power of pattern recognition 

As humans, we use patterns to make sense of our world nearly every day. It’s the basis of our learning — how babies begin to recognize and react to objects. We identify the patterns around us, then associate those patterns with objects or concepts, and respond accordingly based on what we’ve learned from experience.

Brainware approaches data capture in a similar fashion — through pattern recognition.

When it looks at a document, it doesn’t look for specific positions or zones that data are in as a template-based solution would. Rather, it considers all the data on a given document and looks for patterns of information on a page as well as relationships between words.

From the patterns that emerge, Brainware learns what uniquely classifies a document type and which data it should extract. By seeing where clusters of tabular data lie, it can identify whether a document is an invoice vs. a transcript, for example, and then focus on those tables for extraction.

After reviewing just a small set of documents, Brainware understands what patterns classify different document types and then applies that knowledge to classify new documents going forward.

Machine learning enters the room

remote workforce

Hyland also recently began testing its new Automated Learning Engine (ALE) to further enhance Brainware with supervised machine learning capabilities.

What is supervised machine learning? I’m glad you asked.

In the simplest terms, supervised machine learning is when the machine or software learns what to do based on example inputs and outputs provided by humans. When the machine receives new information, it refers back to the training data that has been provided and infers a response.

An early release test version of the ALE is available in the latest release of Brainware for Invoices. The ALE extracts header-level data from invoices and learns from any corrections end users make during the verification process. As users review, verify, or correct extracted data, those corrections are applied to a training set from which the ALE learns to extract the correct data from future invoices.

It’s this combination of pattern recognition and machine learning that makes Brainware and intelligent capture so powerful. In addition to expediting and continuously improving the accuracy of the capture process, the advanced learning capabilities allow users to extract data from all kinds of information sources and handle real-world variation in a way that traditional capture tools simply can’t.

Learn more about the intelligent algorithms powering Brainware here.

Danielle Simer

Danielle Simer is a marketing portfolio manager at Hyland. Her mission is to share best practices and evangelize the power of enterprise content management (ECM) as a tool to automate... read more about: Danielle Simer

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.