Detection of Hate Speech Using Improved Deep Learning Techniques
Jeschelle N. Gallardo, Ernest Dylan G. Gloria, Natalie Rose P. Landicho, Hajah T. Sueno · 2023
LSTM was introduced to solve the problem that RNNs faced, such as suffering from the vanishing gradient problem. Researchers have improved the LSTM model using various techniques. However, most improved models are still based on LSTM’s network structure, which needs to address the weight parameters falling into local extrema. Moreover, they can suffer from poor learning efficiency and gradient disappearance. The paper presented an advanced deep learning methodology aimed at significantly improving the detection of hate speech using LSTM supported by feature extraction, feature selection, and RNN. Our model exhibits outstanding performance, achieving high precision, recall, F1-score, and accuracy at 97%. In comparison, the other models, such as RNN, achieved 56%, LSTM+RNN achieved 80%, By employing RNN with feature extraction and feature selection, an accuracy of 96% was achieved. Likewise, LSTM with feature extraction and feature selection demonstrated comparable results, reaching 96% accuracy across all performance metrics. This shows that the proposed model surpassed the Plain LSTM model employing deep learning techniques.