We present in this paper an original adaptation of Bayesian networks to symbol recognition problem. More precisely, a descriptor combination method, which enables to improve significantly the recognition rate compared to the recognition rates obtained by each descriptor, is presented. In this perspective, we use a simple Bayesian classifier, called naive Bayes. In fact, probabilistic graphical models, more specifically Bayesian networks, are a simple and intuitive way of probability distribution representation. In order to solve the dimensionality problem, we use a variable selection method. Experimental results, obtained in a supervised learning context and tested on GREC symbol database, are very promising.