A new method of representing structured data in neural network classifiers, known as the Compositional Representation has been proposed [4, 5]. This paper gives a brief description of the method, (in particular where it differs from [4]) and applies it to a word recognition task of a isolated multi-speaker data base. The Compositional Representation represents structured data such as strings or lattices of arbitrary length in an n2 dimensional domain (n being a fixed parameter of the model), thus making it possible to use as input for a neural network with n2 input units. The representation can be trained using standard back-propagation. A training algorithm that allocates separate learning factors for each weight is proposed. It has been found that this method converges several orders of magnitude faster than the usual method with just one learning parameter plus momentum term.