It is necessary to be careful in evaluating these results. The results are only as good as the evaluation function, which is certainly flawed in major ways. The texts are of a specific type, there are only three of them and we have not used all rhetorical relations. The evaluations are especially limited by the fact that there is no account taken of the possibilities for aggregation, embedding etc. in the trees that are produced. Nevertheless the approach looks promising enough that it is a real candidate to be used with the ILEX system. Future work needs to look at improving the characterisation of good trees and if possible introducing more natural crossover/mutation operations.