<p>Binary Particle Swarm Optimization (BPSO) has demonstrated effectiveness in discrete data feature selection; however, its performance diminishes in high-dimensional feature spaces. To address this limitation, we present a novel memetic algorithm called Filtration and Local Search-based Binary Particle Swarm Optimization (FLS-BPSO), designed to enhance efficiency and facilitate rapid feature selection in text classification. The algorithm initiates by constructing an initial feature subset using mutual information as a filter-based method. To mitigate unnecessary computation and reduce time complexity, particles showing no improvement after a specific iteration count are filtered out during the process. Achieving a balance between exploration and exploitation, an inertia weight is dynamically updated, accounting for swarm size. Additionally, to maintain diversity and prevent convergence to local optima, a local search technique is applied to the best remaining particles, considering three feature groups: strong, weak, and confusing features. Experimental evaluations on five diverse datasets compare the performance of FLS-BPSO with standard BPSO, the Simulated Annealing (SA) algorithm, and the Genetic Algorithm (GA). Results demonstrate the superiority of FLS-BPSO in terms of classification performance, expressed in terms of micro-F1 score, high feature space reduction rate, and low computational cost when using three classification techniques (NB, SVM, and LSTM), exhibiting enhanced classification accuracy and robustness across multiple datasets varying between binary and multi-class.</p>