When training data is sparse, more domain knowledge must be incorporated into
the learning algorithm in order to reduce the effective size of the hypothesis
space. This paper builds on previous work in which knowledge about qualitative
monotonicities was formally represented and incorp