Correcting recognition errors in handwritten, vocalised Pitman´s shorthand outlines using a Boltzmann machine

Handwritten shorthand is typically 30%-50% faster than the QWERTY keyboard for recording textual information. Therefore, the automatic online recognition and transcription of Pitman´s handwritten shorthand is a potentially faster method of entering textual material into a computer system. Pitman´s shorthand can be considered as composed of two types of symbol: shortforms and vocalised outlines. The vocalised outlines account for between 50% and 70% of shorthand and are composed of a connected string of simple line segments which represent consonant sounds in the word surrounded by simple dots and dashes to represent the vowel sounds. To achieve automatic recognition of these vocalised outlines it is necessary to segment the consonant line into its basic features, classify each of them, including any surrounding vowel symbols and then determine the implied phonetic structure of the word. The authors consider a method of incorporating knowledge about the structure of the consonant kernel of vocalised shorthand outlines in a neural network to detect and correct errors in the initial classification of the basic pattern features and therefore improve the overall recognition performance. The initial results are encouraging