Toward a generalization of neuro-symbolic recognition: An application to arabic words

In this article, we suggest an automated construction of knowledge based artificial neural networks (KBANN) for the holistic recognition of handwritten Arabic words in limited lexicons. First, ideal samples of the considered lexicon words are submitted to a feature extraction module which describes them using structural primitives. The analysis of these descriptions generates a symbolic knowledge base reflecting a hierarchical classification of the words. The rules are then translated into a multilayer neural network by determining precisely its architecture and initializing its connections with specific values. This construction approach provides the network with theoretical knowledge and reduces the training stage, which remains necessary because of variability in styles and writing conditions. After this empirical training stage using real examples, the network reaches its final topology, which enables it to generalize. The proposed method has been tested on the automated construction of neuro-symbolic classifiers for two Arabic lexicons: literal amounts and city names. We suggest the generalization of this approach to the recognition of handwritten words or characters in different scripts and languages.