Cyberbullying has become a pervasive issue on social media, significantly impacting the well-being of individuals, and decision explainability has become a concerning issue in cyberbullying detection. This study aims to enhance the decision explainability of cyberbullying detection by extracting metadata through textual data analysis, based on contextual factors like gender, race, religion, and sentiment. Utilizing the HateXplain dataset, which provides annotated social media posts, we prepare, restructure, and preprocess the data through key steps such as dataset preparation, cleaning, tokenization, and label encoding. Our comprehensive machine learning-based approach focuses on detecting cyberbullying across sensitive categories like race, religion, gender, sexual orientation, and miscellaneous criteria. We employ a Bi-LSTM model to capture the contextual and sequential dependencies inherent in online harassment. The model’s performance is evaluated using various metrics, including top-K accuracies, precision, recall, and F1-score. The results highlight the model’s effectiveness in addressing the nuanced and context-dependent nature of cyberbullying, offering a well-prepared dataset and a robust solution for fostering safer online environments.