Planar Markovian approach for the recognition of a wide vocabulary of Arabic decomposable words

The recognition of Arabic writing is still an important challenge because of its flexional nature and great topological variability. For that, we have been investigating the use of linguistic knowledge to improve the recognition of wide Arabic word lexicon. In this paper, we propose a hybrid approach for the recognition of decomposable Arabic words by adopting a planar Markovian modeling where the first dimension embodies the morphology of the language and the second is devoted to the topology of the script. Indeed, the proposed model includes 101 planar hidden Markov models (PHMM), each of them is dedicated to learning and recognizing a sub-vocabulary derived from one root. On one hand, each implements the rules of the morphology of the Arabic (derivation, flexion and agglutination). On other hand, each classifier models the topological properties of Arabic letters using global and local primitives. Given that each meta-state and state of the main HMM represents a definite morphological element (root letter, infix, enclitic …), we opted for supervising training while specifying the Viterbi path that must maximize the likelihood. We handled a wide vocabulary of 7022 words got from 101 roots. Experiments were conducted on a corpus of more than 21000 samples and yielded promising results (top2 = 92.37%).