Exploiting qualitative domain knowledge for learning Bayesian network parameters with incomplete data