The ever-increasing use of social media in the internet space have induced a number of problems like cyberbullying and cyberaggression over the internet. Researchers have made a commendable progress on the ongoing fight against cyberbullying but a lot of unresolved issues still persist that primarily motivates the purpose of the research. The paper aims to integrate recent advances in the field of word embedding like fastText, ELMo and stacked flair embeddings combined with a host of robust deep learning techniques to further the efficiency of detection over the state-of-art. Two distinct datasets Formspring and Wikipedia were requested and processed for the purpose of the research. A number of different combinations of word embedding with deep learning methods were tested and compared with CNN with ELMo embedding delivering the most promising results with an F1 score of 0.82 on both datasets. On the other hand, CNN with fastText obtained F1 score of 0.82 on Formspring and 0.64 on Wikipedia dataset but was computationally faster than the counterparts. Moreover, transfer learning was performed using the models to test and prove the robustness and efficacy of the models. The system performed considerably well with superior scores in precision, recall and F1 over the state-of-the-art across all the test cases performed.