We propose to learn the parameters of evidential contextual correction mechanisms from a learning set composed of soft labelled data, that is data where the true class of each object in only partially known. The method consists in optimizing a measure of discrepancy between the values of the corrected contour function and the ground truth also represented by a contour function. The obtained values are then used to correct the output of the classifier. The advantages of this method are illustrated by tests we have done on synthetic and real data.