Neural-linguistic classifier combination for large Arabic word vocabulary recognition

A new system for large Arabic word vocabulary recognition is proposed. It is based on combination of three neural-linguistic classifiers that collaborate to improve the recognition results. Our work confirms not only “word superiority effect”, which is a concept in human reading process, but also proposes a new concept named “word derivation effect”. In fact, the proposed classifiers are based on transparent neural networks (TNNs) which use global features such as ascenders, descenders and loops. Thus, the system proceeds firstly by global structural primitives and then refines the vision by looking for local details, in cases of ambiguities. Inspired by previous works, dealing with the use of Arabic linguistic knowledge in writing recognition, we investigate the use of a linguistic based collaboration between TNNs to handle with a large Arabic word lexicon that involves decomposable words derived from healthy roots. Indeed, our neural classifiers, called TNN_R, TNN_S and TNN_C, are conceived to collaborate and respectively learn and recognize roots, schemes and conjugation elements of words. Comparisons between results of “classifier combination without collaboration”, results of “classifier collaboration” and results of “applying perceptive cycles” highlight the contribution of collaboration and perceptive cycles. The proposal system has been successfully demonstrated on a vocabulary of 2000 words and achieved satisfactory results compared to some related works, which either deal with large vocabulary and/or integrate linguistic knowledge in their system.