Adaptive multilingual cyberbullying detection via grammatical evolution

The rapid expansion of digital communication has made social media an integral part of daily life, yet it has simultaneously amplified the reach of harmful behaviors. Early detection of cyberbullying is recognized as a critical safeguard for mental health, particularly for vulnerable demographics. Deep learning models, specifically multilingual transformers, are currently widely used as the most effective tools for automating toxicity detection across diverse languages. However, the interpretation of multilingual abuse is a complex multi-parameter process that frequently leads to false-positive and false-negative results. Static architectures often fail to capture evolving adversarial patterns, such as code-switching and dialectal sarcasm (e.g., the ironic use of حلوة أوي “very sweet”), and frequently reproduce algorithmic biases against protected groups. This article introduces a novel adaptive framework that synergizes Grammatical Evolution (GE) with hybrid multilingual embeddings to classify cyberbullying content with high precision and fairness. Distinct from traditional fixed architectures, our approach utilizes GE as an evolutionary optimizer to autonomously search the topological space of bidirectional long short-term memory (BiLSTM) classifiers, tailoring neural structures to the semantic richness of hybrid XLM-R and multilingual bidirectional encoder representations from transformers (mBERT) representations. Furthermore, a fairness-constrained fitness function is integrated to explicitly mitigate algorithmic bias during the optimization process. Experimental evaluation on the ArCyC Arabic corpus and the Fine-Grained Balanced English dataset demonstrates that this evolutionary-deep learning hybrid achieves a state-of-the-art F1-score of 0.9458. Crucially, the framework reduces subgroup performance disparities across age, gender, and religion by 41%, validating that ethical alignment can be mathematically integrated into the learning process without compromising predictive accuracy.