Realising a general purpose artificial intelligence is an ambitious goal. To do this, new techniques, especially in an open-world context where new tasks may appear, are to be developed to deal with the dynamism and diversity of these systems. A relevant example of this kind is Zero-Shot Learning, where training and test sets have disjoint label sets, so that, no examples of the classes expected in the test (unseen classes) are available during the training phase, and the training (seen) classes are not involved in the test phase. Recent studies aim to generalise from seen to unseen classes by building a variety of models that exploit semantic information in addition to the actual samples (e.g. images). However, the adequacy of the input semantic attributes has not been explored. This initial work studies the influence of preprocessing the semantic space using embedded feature selection within a cross-validation scheme to generalise to unseen classes. The experiments on three well-known zero-shot learning benchmarks show that selecting a subset of the semantic attributes may be beneficial to learning more general methods.