<![CDATA[The rapid growth of social media platforms has intensified the need for accurate cyberbullying detection systems capable of understanding contextual and linguistically complex expressions. Existing machine learning and deep learning approaches often suffer from limited interpretability, insufficient contextual understanding, and suboptimal parameter optimization, reducing their effectiveness in identifying harmful online content. This study proposes a novel cyberbullying detection framework that integrates Linguistic Rule-Based Feature Extraction, an Elman Neural Network (ENN), and the Local Search-Based Improved Bat Algorithm (LSBIA). The main contribution of this research lies in the synergistic combination of interpretable linguistic knowledge, contextual sequence modeling, and metaheuristic optimization within a unified classification framework. Linguistic rules are employed to capture negation patterns, intensifiers, and adjective–noun relationships, while ENN models contextual dependencies through recurrent memory structures. LSBIA is utilized to optimize network parameters and improve convergence stability. Experiments were conducted using textual data collected from Instagram, Twitter, and Facebook and evaluated using stratified 10-fold cross-validation. The proposed method achieved an accuracy of 99.12%, precision of 94.73%, recall of 97.45%, and F1-score of 93.91%, outperforming Support Vector Machine (91.20% accuracy), Naïve Bayes (89.75%), and Decision Tree (90.10%). Ablation experiments further demonstrated the importance of each component, where removing linguistic rules reduced accuracy to 94.90%, removing sentiment scoring reduced accuracy to 96.30%, and replacing ENN with LSTM, GRU, or Transformer architectures resulted in lower accuracies of 92.50%, 91.90%, and 93.20%, respectively. These findings confirm that integrating linguistic feature engineering, contextual neural modeling, and metaheuristic optimization significantly enhances cyberbullying detection performance while maintaining interpretability. The novelty of this study resides in the integration of linguistic rule-based representation with LSBIA-optimized ENN for context-aware cyberbullying classification.]]>
