Cyberbullying Detection: Hybrid Models Based on Machine Learning and Natural Language Processing Techniques

The rise in web and social media interactions has resulted the efortless proliferation of offensive language hate speech. Such online harassment, insults, attacks are commonly termed cyberbullying. sheer volume user-generated content made it challenging to identify such illicit content. Machine learning wide applications text classification, researchers shifting towards using deep neural networks detecting cyberbullying due several advantages they have over traditional machine algorithms. This paper proposes a novel network framework with parameter optimization an algorithmic comparative study eleven classification methods: four seven shallow on two real world datasets. In addition, this also examines effect feature extraction word-embedding-techniques-based natural processing performance. Key observations from show that bidirectional attention models provide high results. Logistic Regression was observed be best among classifiers used. Term Frequency-Inverse Document Frequency (TF-IDF) demonstrates consistently accuracies techniques. Global Vectors (GloVe) perform better models. Bi-GRU Bi-LSTM worked amongst extensive experiments performed datasets establish importance work by comparing methods Our proposed outperform existing state-of-the-art approaches for detection, accuracy F1-scores as ~95% ~98%, respectively.