Comparison of Deep Learning Methods in Detecting Hate Speech in Indonesian Tweets

Hate speech has negative effects on both the targeted victims and the listeners. The dissemination of hate speech can occur not only physically or verbally, but also in writing on social media. The emergence of hate speech on social media platforms can be difficult to identify in written communication. Currently, hate speech detection relies on machine learning. This study generates a vector representation of words using three pre-trained word insertion models: Global Vectors (GloVe), FastText, and Bidirectional Encoder Representations from Transformers (BERT). Synthetic Minority Oversampling Technique (SMOTE) and Random Over Sampling (ROS) were utilized as balancing methods to rectify data imbalance between classes. In addition, three distinct deep learning architectures were used to identify sentence-level hate speech in Indonesian tweets: Bidirectional Long Sort-Term Memory (BiLSTM), Convolution Neural Network (CNN), and Recurrent Neural Network (RNN). The dataset was collected by crawling the data via the Twitter API. After data underwent preprocessing, characteristics were extracted. Based on experimental results, classifiers employing RNN and BERT embedding and utilizing SMOTE produced the most accurate results (95.5%).