Abstract: We disclose a trainable recognizer for multi-stroke symbols. Pen speed and the curvature are examined to segment a stroke into the constituent primitives. The primitives are determined to be lines or arcs using, for example, a least squares best fit method. Symbol recognition (training) results in the construction of definitions consisting of the constituent geometric primitives (lines and arcs), the properties of individual primitives, and the geometric relationships between them. A definition is learned by examining a few examples of a symbol and identifying which properties and relationships occur frequently. Thereafter, during symbol recognition (matching), an unknown symbol is identified by determining which definition matches the unknown with the least error. We disclose two recognition methods. One assumes that the primitives of a symbol are always drawn in the same order. This method is fast, but requires some care from the drawer.