Abstract: A character-recognition unit or other input device provides a sequence of input characters. A questionable identification of one of the input characters is resolved by considering surrounding characters by a modified N-gram technique. The context characters are each translated into one of a small number of reduced codes each representing a subset of the alphabet to be recognized. The reduced codes are combined to produce a context code which indicates a set of categories containing the correct identification for the questionable input character. A list of confusion codes specifies a number of possibly correct identifications for one of the input characters. One of the codes in the confusion list is selected by the context code as the correct identification for the questionable character.