Estimating residual error rate in recognized handwritten documents using artificial error injection

Both handwriting recognition systems and their users are error prone. Handwriting recognizers make recognition errors, and users may miss those errors when verifying output. As a result, it is common for recognized documents to contain residual errors. Unfortunately, in some application domains (e.g. health informatics), tolerance for residual errors in recognized handwriting may be very low, and a desire might exist to maximize user accuracy during verification. In this paper, we present a technique that allows us to measure the performance of a user verifying recognizer output. We inject artificial errors into a set of recognized handwritten forms and show that the rate of injected errors and recognition errors caught is highly correlated in real time. Systems supporting user verification can make use of this measure of user accuracy in a variety of ways. For example, they can force users to slow down or can highlight injected errors that were missed, thus encouraging users to take ...

• Type: Conference
• Year: 2010
• Where: CHI
• Authors: Edward Lank, Ryan Stedman, Michael Terry
• Keywords: CHI 2010, Human Computer Interaction, Injected Errors, Recognition Errors, Residual Error