Experimental comparison of algorithms for symbols’ classification in shorthand documents

A relative position of the symbol influences its meanings. This dependence was revealed in the process of decoding historic shorthand written documents. There are three positions: basic, superscript, or subscript. We present a comparison of the four algorithms for symbols’ classification based on the single approximation method, double approximation method, single approximation method with the error decrease, and the double approximation method with the error decrease. Parameters of the algorithms were chosen experimentally using a validation set. The performance level of the algorithms was measured in terms of accuracy, precision, recall, F-measure, and summarized F-measure. Based on the summarized F-measure, the best result was achieved with the algorithm for symbols’ classification by double approximation method with the error decrease. We identified parameters for each algorithm in which the summarized F-measure is maximized for the validation data. We also identified learning duration period for each algorithm.