A classification system called ECHO has been designed for our participation to the first LSHTC Pascal challenge. This system is based on a principle of echo. It computes the score of a document for a class by combining a bottom-up and top-down propagation of activation in a very simple neural network. For this second edition of the challenge, the system has been refined in order to improve its performances. The main improvement concerns the definition of a method of calibration integrating the distribution of the echo scores of the training documents for each class. Another improvement concerns the management of the relative importance of bottom-up and top-down propagation which we make vary with the class.