This paper presents a novel approach to the classifier selection problem with text data, leveraging efficient dataset representations and evolutionary hyper-heuristics. A hyper-heuristic is modeled as a set of blocks, with each block as a meta-feature vector labeled with a classifier. The initial solution space consists of 16 meta-features. For a dataset, a hyper-heuristic selects the nearest block to its meta-feature vector, returning the associated classifier. A genetic algorithm evolves the hyper-heuristics through tailored evolutionary operators, and their performance is evaluated across a group of datasets using the macro F1 score. After evolution, the best hyper-heuristic is tested on unseen datasets to evaluate generalization. Principal Component Analysis is applied to select the most relevant meta-features based on explained variance, reducing the space to eight features in which the hyper-heuristics are re-evaluated. The best hyper-heuristics from both spaces are compared to validate the reduction technique. The results show similarities in classifier labels, performance, and computational cost. Finally, the proposed approach is compared with four state-of-the-art automated machine learning systems, AutoKeras, AutoGluon, H2O, and TPOT, demonstrating competitive classification performance with lower computational requirements.