An important problem of artificial intelligence is learning class description from pre-classified examples. The emphasis of some of the important learning systems such as ID3, INDUCE and CART is to discriminate each class from every other class. In many practical cases such descriptions are very inappropriate. In this paper, we describe a learning system that uses a reference description to learn each class description. The use of the reference description ensures learning of a class description that describes the class in addition to discriminating it from all other classes. Moreover, the description of each class is such that characterising attributes are specified before discriminating attributes. This is a major advantage over an earlier learning system called KAHLE. The reference and class descriptions learnt are shown to converge in the stochastic sense. The class description thus generated is simplified by dropping attributes which do not add to the description in any way. The importance of an attribute for a class is determined from this description. This is used in inference of a test example with missing attribute values. An inference process using the importance of attributes and based on category validity is used to classify test examples. The problem of characterisation of a democrat and a republican based on the machine learning database maintained at the University of California, Irvine is handled well by the proposed system. The results demonstrate that a better description is not at the expense of classification accuracy.