The most common document formalisation for text classi?cation is the vector space model founded on the bag of words/phrases representation. The main advantage of the vector space model is that it can readily be employed by classi?cation - gorithms. However, the bag of words/phrases representation is suited to capturing only word/phrase frequency; structural and semantic information is ignored. It has been established that structural information plays an important role in classi?cation accuracy [14]. An alternative to the bag of words/phrases representation is a graph based rep- sentation, which intuitively possesses much more expressive power. However, this representation introduces an additional level of complexity in that the calculation of the similarity between two graphs is signi?cantly more computationally expensive than between two vectors (see for example [16]). Some work (see for example [12]) has been done on hybrid representations to capture both structural elements (- ing the graph model) and signi?cant features using the vector model. However the computational resources required to process this hybrid model are still extensive.