Verifying the UNIPEN devset

This paper describes a semi-automated procedure for the verification of a large human-labeled data set containing online handwriting. A number of classifiers trained on the UNIPEN "trainset" is employed for detecting anomalies in the labels of the UNIPEN "devset". Multiple classifiers with different feature sets are used to increase the robustness of the automated procedure and to ensure that the number of false accepts is kept to a minimum. The rejected samples are manually categorized into four classes: (i) recoverable segmentation errors, (ii) incorrect (recoverable) labels, (iii) well-segmented but ambiguous cases and (iv) unrecoverable segments that should be removed. As a result of the verification procedure, a well-labeled data set is currently being generated, which will be made available to the handwriting recognition community.