Neural nets for real time analysis of cursive writing

Abstract

The display of Arabic in videotext systems poses special problems because Arabic characters change shape depending on their context in a word. Thus, it is necessary to make a contextual analysis before the Arabic characters are displayed. This analysis can cause a loss of characters in some constraining conditions of display time. This is the case with the videotext service governed by the Télétel Standard norm (Teletel, 1986; Saboureau and Bouche, 1984, Guide pratique du videotex et du Minitel, CEDIC edition, Paris; Anne and Thiebault, 1990, Guide pratique du videotex, Eyrolles edition, Paris; Matri et al., 1990, Télématique, Techniques, Normes et services, Dunod edition, Paris), which makes use of the standard V23 for the transmission of characters (1200 bits/s 75 bits/s). Therefore, it is important to reduce the contextual analysis duration in display systems. In order to optimize this duration, we have elaborated upon a display system in which the architecture is based on the interconnection of neural networks (MacCulloch and Pitts, 1943, A logical calculus of the ideas immanent in nervous activity, Bulletin of Mathematical Biophysics 5, 115–133; Hebb, 1949, The organization of behaviour, Wiley, New York; Rosenblatt, 1958, The perceptron: a probabilistic model for information storage and organization in the brain, Psychological Review 65, 386–408; Rumelhart et al., 1986a, Learning internal representations by error propagation, in: Parallel Distributed Processing: Exploration in the Microstructure of Cognition, Vol. 1, MIT Press, Cambridge, MA). The latter have shown its effectiveness in may problems in real time such as the recognition of speed, artificial vision and motor control. Our system contains three neural multilayer nets (Rosenblatt, 1958, The perceptron: a probabilistic model for information storage and organization in the brain, Psychological Review 65, 386–408; Rumelhart et al., 1986a, Learning internal representations by error propagation, in: Parallel Distributed Processing: Exploration in the Microstructure of Cognition, Vol. 1, MIT Press, Cambridge, MA; Rumelhart et al., 1986b, Learning representations by back-propagating error, Nature 323, 533–536). The learning of the neural net uses the classic algorithm of error backpropagation (Rumelhart et al., 1986a, Learning internal representations by error propagation, in: Parallel Distributed Processing: Exploration in the Microstructure of Cognition, Vol. 1, MIT Press, Cambridge, MA; Rumelhart et al., 1986b, Learning representations by back-propagating error, Nature 323, 533–536; Pao, 1989, Adaptive pattern recognition and neural networks, Addison-Wesley, New York). When we used this system in a bilingual videotext emulator (Arabic/Latin) it yielded satisfactory results. However, it possesses the inconvenience of being specific to one set of Arabic characters.