Cyberbullying detection using deep learning and word embeddings: an empirical study

Cyberbullying detection has become a pressing need in Internet usage governance due to its harmful consequences. Different approaches have been proposed to tackle this problem, including deep learning. In this paper, an empirical study is conducted to evaluate the effectiveness and efficiency of deep learning algorithms, coupled with word embeddings in detecting cyberbullying texts. Three deep learning algorithms were experimented, namely GRU, LSTM and BLSTM. Data pre-processing steps, including oversampling were performed on the selected social media datasets. For feature representations, four different word embeddings models were explored, including word2vec, GloVe, Reddit and ELMO models. Elmo cares of word context by capturing information from the word surroundings which eliminates some of the shortcomings of pre-trained word embeddings models. For more accurate results, 10-fold cross-validation technique was implemented. The experimental results show that BLSTM performs best with ELMO in detecting cyberbullying texts. The efficiency of each model is also measured by calculating the average time taken for training each model. GRU outperforms in terms of time efficiency. Based on the analysis done on false negative cases, three observations were made, which highlight the limitations of word embeddings models on top of GRU algorithm in cyberbullying detection.