The use of social media and online communication has made cyberbullying a serious issue. The large amount of user-generated content makes manual monitoring difficult, creating the need for accurate automatic detection systems. Nevertheless, most of the current approaches are limited to binary classification and fail to integrate semantic and linguistic data. This paper suggests a multi-step model of cyberbullying detection using machine learning and deep learning methods. The suggested framework integrates sentence embeddings produced by the Sentence-BERT model with manually designed linguistic features of aggression, threats, punctuation, and writing style. Three successive stages are investigated. The initial step uses classical machine learning models, such as Support Vector Machine, Random Forest, and Logistic Regression. The second phase presents hybrid deep learning models that are founded on dense and convolutional neural networks. The third phase builds upon the framework with a cross-attention mechanism and a stacking meta-ensemble classifier to enhance feature interaction and prediction accuracy. The experiments were performed on the Cyberbullying Tweets dataset with binary and multi-class labels. The binary setting is used to differentiate between cyberbullying and non-cyberbullying, whereas the multi-class setting is used to distinguish between various types of cyberbullying. The proposed Stage 3 framework achieved the best performance, reaching 91.0% accuracy and 0.90 F1-score in binary classification, and 88.0% accuracy and 0.875 F1-score in multi-class classification. The results demonstrate that combining semantic representations, linguistic features, attention mechanisms, and ensemble learning provides a more effective and reliable solution for cyberbullying detection than traditional approaches.