Hindi Educational Text Classification Using Convergent PSO-Optimized Heterogeneous Ensemble Framework with IndicBERT Embeddings

Abstract

This work proposes a Convergent PSO-Optimized Heterogeneous Ensemble Framework for Hindi educational text classification, addressing the challenges of high-dimensional embedding spaces and class imbalance in low-resource languages. The framework integrates Particle Swarm Optimization (PSO) with IndicBERT embeddings, ensuring swarm stability through a metric-geometric formulation that constrains particle dynamics by Lipschitz conditions in the embedding manifold. Moreover, synthetic minority-class samples are generated directly in the embedding space using SMOTE, preserving semantic structure while balancing class distributions. The heterogeneous ensemble combines a Hierarchical Attention Network, a Kernelized SVM, and LightGBM, whose hyperparameters are dynamically optimized by PSO to maximize macro-F1 performance. The proposed method uniquely unites geometric swarm stability, embedding-space augmentation, and ensemble tuning, offering a robust solution for Hindi text classification. Experimental validation demonstrates significant improvements over conventional approaches, particularly in handling imbalanced educational datasets. This work contributes a theoretically grounded and empirically validated framework for low-resource language processing, with potential applications in multilingual educational content analysis.