The filtered combination of the weighted edit distance and the Jaro-Winkler distance to improve spellchecking Arabic texts

The digital environments for human learning have been much evolving thanks to the incredible progress of information technologies. This is particularly the case for automatic correction of spelling errors requested by a large majority of people nowadays. Almost all of the current spellcheckers are semiautomatic, and they enable users to find the good solution for a committed error. The major shortcoming of the existing metric methods of correction lies in the bad scheduling of the solutions suggested to the spellchecking out of context of a detected error. To overcome this limitation, we have developed several approaches which suggest probability costs estimated from a learning test. It is attributed in various editing operations during calculating measure of similarity, case of the edit distance. The idea developed in this work was to know how to efficiently weigh these editing operations without resorting to a phase of learning. This is based only on the proximity and the similarity between Arabic keyboard keys. Additionally, we have suggested combining this measure with the distance of Jaro-Winkler in order to better filter, refine and weigh certain solutions compared to others. The experimental results stem from tests conducted on errors committed in a learning corpus, trying to validate the choices of conception and to prove the interest of both approaches.