Harassment in schools and workplaces causes severe effects like anxiety, depression, decreased self-esteem, and even suicide. While school harassment includes physical and psychological bullying, workplace harassment involves hostility and manipulation. With the rise of the internet and social media, harassment has increasingly moved to digital platforms, exacerbating its impact due to the persistence of these media. This study addresses a critical gap in the existing literature by employing multiple advanced deep learning techniques for classifying harassment texts on social media, each evaluated independently. Unlike traditional methods, our approach utilizes diverse models including dense neural networks, CNNs, LSTMs, and Graph Convolutional Networks to provide a comprehensive analysis of classification performance. By focusing on the unique challenges of cyberbullying detection, this work contributes evidence-based strategies for early detection and effective intervention. Notably, the combination of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) achieved the highest precision of 0.840 and an F \(_1\) score of 0.836. In contrast, the Autoencoder and k-means method exhibited lower performance, with a precision of only 0.457 and an F \(_1\) score of 0.435.