LSTM-BASED MODEL FOR CYBERBULLY DETECTION

Abstract

Cyberbully has become rampant due to digitalization and conventional cyberbully detection is time consuming, this led to the development of cyberbully detection systems. Previous  cyberbully detection systems yielded low accuracy, hence, this research developed a LSTM-based model for cyberbully detection. The dataset for training the model was obtained from Kaggle and pre-processed by removing punctuation marks and stop words, stemming, tokenization and one hot representation. Feature extraction was done on the datasets to remove outliers and Python 3.9 was used for implementation. The developed system was evaluated using: accuracy, precision, Recall and F1 measure and the results obtained were compared to other machine learning models as well as a hybrid of CNN-LSTM. Result shows The developed model yielded an accuracy of 77.0% with a validation time of 3.024 sec in the detection of cyberbully while the hybridization of LSTM-CNN gave an accuracy of 74.80% for fake news and cyberbully detection. The developed model was also bench marked with other machine learning models: SVM, KNN and RF and the system developed outperformed them. The outcome of this research show that deep learning approach used outperformed the machine learning models considered in this research for cyberbully detection. However, future research should employ locally collected dataset for cyberbully detection.