- The holistic approach for word recognition has become widely accepted in Arabic text recognition research. However, the large search space limits this approach to domains with small vocabularies. In this work, we present a novel approach to generate hierarchical representation for the shapes of Arabic continuous sub-words. The top levels of the hierarchy include the coarse representations of sub-words, and the low levels include the fine representations. The construction of the hierarchy is performed bottom up; the shapes at each level are simplified and classified to generate the next level. The search for an appropriate match is performed top down; at each level it traverses sub-trees, whose roots have the highest match rate.