An essential ingredient: Post recognition processing By Reynolds C. Bish, President TextWare Corporation Conventional wisdom tells us that forms processing is easy. You've heard the story. Scan the documents, use recognition technologies to automate the process, get rid of your key entry operators and slash costs. It sounds simple... but it's not. No recognition technology provides a 100% solution. Imaging applications can be broadly separated into two categories: document management and forms processing. Document management applications focus on retrieving and managing the distribution of images. Erroneous index data can be tolerated by using multiple indexes per document and through the use of "fuzzy" searching algorithms. Forms processing applications focus on capturing data from a form. Images are a vehicle to enable more efficient capture of that data. The cost of correcting erroneous data in mainstream data processing systems can be enormous and, therefore, accuracy is essential. What makes forms processing applications more difficult to implement and prone to producing erroneous data? Unfortunately, no two forms or scanned images of the same form are identical. Many variables in the design, production, completion and scanning of forms can adversely affect scanning and recognition technologies. In addition, recognition accuracy varies greatly and is highly dependent upon the type of data being processed. High-quality machine print typically results in accuracy rates that approach 100%. Conversely, low-quality, unconstrained hand print produces much lower rates--as low as 10% in some cases. Cursive data inevitably produces meaningless results. Another major problem is that quoted accuracy rates can be very misleading. First, while tests typically report 97% or better accuracy rates, it's important to remember that those tests use full-text documents with high-quality machine print, not the real-world forms we all contend with on a daily basis. Second, we need to understand what this number really means. In fact, it's a "read rate." The recognition engine only knows it didn't recognize 3% of the characters. The other 97% may or may not be correct. It thinks they're correct but doesn't really know. It ignores a number of very real potential issues--substitutions, the effects of excess or insufficient characters, data collected from the wrong area of the form, accurately recognized but still invalid data and many other problems. Benefit of Post-Recognition Processing Finally, quoted accuracy rates are reported as a percentage of characters processed, not as a percentage of the fields in a document. The latter statistic is more relevant in forms processing applications. The reason is obvious. If there's an error in a field, a key entry operator needs to access that field and manually repair one or more characters. Productivity suffers, costs escalate and the savings from automating data entry can evaporate. Field level accuracy rates are always significantly lower than character level accuracy rates. Assuming an equal distribution of unrecognized characters in all fields, a field level accuracy rate can be calculated by taking the character level accuracy rate to the power of the number of characters in a particular field. Character level accuracy rates may be above 80% and field level accuracy rates can still be as low as only 20%. And, remember, these are quoted rather than true accuracy rates -- which, as we know, are likely to be even lower. All these factors combine to prove that no recognition technology provides a 100% solution. Post recognition processing is essential to ensure the accuracy of all data captured. The Importance of Data Perfection and Completion Post-recognition processing takes place after an image has been pushed through a recognition engine but before the resulting data is sent to a mainstream data processing system. It serves two functions. First, it provides what we might call "data perfection." That is, it ensures the accuracy of data captured via recognition engines. Second, it enables what we might call "data completion." In other words, it lets us capture data that can't be pushed through recognition engines. How is this accomplished? Post recognition processing starts with an automated review of the data produced by recognition engines. Application-specific programs are used to identify unrecognized characters, suspicious characters, outright errors, substitutions, fields with too many or too few characters and data that may have been accurately recognized but is, in fact, invalid. In other words, the post processing focuses on the shortcomings of recognition technologies. Field vs. Character Level Accuracy It then attempts to automatically correct these shortcomings and uses valid data to retrieve and insert related data. For example, a zip code may be used to retrieve and insert city and state information, and all three may then be used to validate or correct the related name and street address. These application-specific programs may utilize contextual analysis, dictionaries, tables, databases, standard or custom algorithms and many other data validation routines and techniques. The goal is to minimize the amount of manual effort required to subsequently produce accurate and complete data. The next step requires three different groups of key entry operators. The first group handles OCR/ICR repair to key unrecognized characters, including accepting or correcting suspicious characters. If a field contains too many unrecognized or suspicious characters, operators should be forced to key the entire field, since this is more productive than repairing individual characters. The second group keys fields that contain invalid data or data that can't be pushed through recognition engines. Keying data from images is very different from OCR/ICR repair because complete fields are always keyed. As a result, operators can anticipate the next field and develop more of a rhythm. The third group "blindly" keys all data in critical fields and simultaneously compares the characters being keyed to the data previously captured. This is referred to as verification and is the only way to ensure 100% accuracy. The objective is for all three groups to complete their work as quickly, efficiently and accurately as possible. Therefore, high-speed data entry capabilities and tools to measure and manage operator productivity are required. Separating the tasks significantly increases productivity, and one pass--in context approach with simultaneous, automatic validation of the characters being keyed--is essential. The last step in post recognition processing ensures the overall accuracy and validity of all data. This is accomplished with an automated review using application-specific programs that validate accuracy on a character, field, form and batch basis. For example, an accumulator may be used to add the amount in a field on each form and compare that total to a separate batch balance amount. Characters, fields, forms or entire batches that fail need to be flagged for further exception processing. The data and the images can then be output or uploaded to the user's mainstream data processing system. As we've seen, post-recognition processing requires several steps and different groups of operators. As a result, the entire process needs to be batch-oriented and workflow-controlled to complete these steps as quickly and efficiently as possible. The Benefit of Post- Recognition Processing So, have we totally negated the benefits of recognition technologies in forms processing applications? Absolutely not. Manual data entry is expensive -- more than $22 per hour according to published surveys. Automating this process with recognition technologies can eliminate much of this work. Post recognition processing allows us to perfect and complete the data. Together they produce enormous benefits. To illustrate this point, TextWare has developed a concept called "equivalent keystrokes per hour," which combines the effectiveness of recognition technologies and post recognition processing. According to published surveys, a key entry operator averages 11,263 keystrokes per hour. The combination of recognition technologies and post-recognition processing produces a higher throughput rate, even with true accuracy rates as low as 50%. With true accuracy rates above 80%, the advantage becomes enormous. Let's look at a real-world example. TextWare has a customer that processes hundreds of thousands of health insurance claims every day. Each claim contains approximately 300 characters to be captured. If they were to key and verify all the data, it would cost $1.20 to process each form. By applying recognition technologies and our post-recognition processing software, they've reduced this by 17% to $1.00. And they're capturing fewer than 50% of the characters via recognition technologies, realizing only an 84% true accuracy rate and still verifying all the data. Obviously, with effective post recognition processing, users don't need to capture all data via recognition technologies and true accuracy rates as high as 97% to justify the cost of forms processing applications. This is the most important but often-overlooked benefit of post recognition processing -- it significantly increases the number of viable forms processing applications. It is, indeed, an essential ingredient. * Reynolds C. Bish is President of TextWare Corporation (Park City, UT), a leading supplier of software for forms processing, post recognition processing and image assisted data entry applications on client/server platforms. The company markets its products directly to large end users and through systems integrators and international distributors. He can be reached at: Phone 801-645-9600 Fax 801-645-9610 E-mail bish@textware.com IW Special Supplement, March 1996
	TOP OF PAGE		HOME ++ SEARCH IW ++ DAILY NEWSWIRE ++ CURRENT PUBLICATIONS ++ ABOUT IW BUYER'S GUIDE ++ DIIME ++ SUBSCRIBE ++ COMMENTS
	© 1995, 1996, 1997 Cardinal Business Media, Inc. All Rights Reserved. The names, logos and icons identifying CBMÆs products and services are proprietary marks of Cardinal Business Media, Inc. CBM has no liability for content or goods on the Internet except as set forth in the Terms and Conditions of Service.