HMM Based Model for Arabic Word Conjugation Recognition and Word Reconstitution from Morphological Concepts

摘要

In this paper we present a new Markovian approach for the training and the recognition of the conjugation (e.g number, gender, tense) of a large vocabulary of Arabic decomposable words. The recognition is done from global structural primitives (e.g ascender, descender, loop, diacritics). Our system is based on 12 Hidden Markov Models (HMMs), each one is specific in one inflectional conjugation element (e.g feminine, masculine, singular, dual, plural, present, past), to help classification in a vocabulary of 5757 words. Our HMMs architectures integrate linguistic properties of inflectional conjugation of an Arabic radical. This work is part of a system conceived to recognize words within the recognition of their morphological concepts (root, schemes and conjugations). Experiments, conducted on a corpus of 17357 words images, gave encouraging rates.