In recent years, most content-based spam filters have been implemented using Machine Learning (ML) approaches by means of token-based representations of textual contents. After introducing multiple performance enhancements, the impact has been virtually irrelevant. Recent studies have introduced synset-based content representations as a reliable way to improve classification, as well as different forms to take advantage of semantic information to address problems, such as dimensionality reduction. These preliminary solutions present some limitations and enforce simplifications that must be gradually redefined in order to obtain significant improvements in spam content filtering. This study addresses the problem of feature reduction by introducing a new semantic-based proposal (SDRS) that avoids losing knowledge (lossless). Synset-features can be semantically grouped by taking advantage of taxonomic relations (mainly hypernyms) provided by BabelNet ontological dictionary (e.g. “Viagra” and “Cialis” can be summarized into the single features “anti-impotence drug”, “drug” or “chemical substance” depending on the generalization of 1, 2 or 3 levels). In order to decide how many levels should be used to generalize each synset of a dataset, our proposal takes advantage of Multi-Objective Evolutionary Algorithms (MOEA) and particularly, of the Non-dominated Sorting Genetic Algorithm (NSGA-II). We have compared the performance achieved by a Naïve Bayes classifier, using both token-based and synset-based dataset representations, with and without executing dimensional reductions. As a result, our lossless semantic reduction strategy was able to find optimal semantic-based feature grouping strategies for the input texts, leading to a better performance of Naïve Bayes classifiers.