Illuminating the Shadow: Building an Advanced Semantic Cyberbullying Detection System

Cyberbullying on social media has become a pervasive and deeply concerning issue in today's digital age. With the ease of anonymity and the widespread accessibility of online platforms, individuals of all ages are increasingly vulnerable to harassment, intimidation, and cruelty. The impact on victims can be devastating, leading to profound psychological distress, anxiety, depression, and even thoughts of self-harm or suicide. This paper proposes an automatic cyberbullying detection system for online comments. To achieve this objective, we followed four main steps: data collection and preprocessing, semantic processing, classification, and evaluation. We combined WordNet with GloVe word embeddings to cover lexico-semantic categories in comments and capture the semantic relationships between synonyms. In addition, we tested three synonym replacement strategies during semantic processing: (1) replacing a word with its closest synonym, (2) Type-1 fuzzy synonym replacement, and (3) Type-2 fuzzy synonym replacement. For the classification, we used Logistic Regression, Random Forest, Decision Tree, Extremely Randomized Trees, XGBoost, AdaBoost, LSTM and RNN in order to build robust models capable of accurately detecting cyberbullying behaviors, thereby offering an effective solution to promote a safer and more benevolent online environment. Experimental results show that our proposed system was able to detect cyberbullying with an Accuracy of 94.36%.