Sara Detection on Social Media Using Deep Learning Algorithm Development

Abstract

Social media has become a key platform for disseminating information and opinions, particularly in Indonesia, where SARA (Ethnicity, Religion, Race, and Intergroup) issues can fuel social tensions. To address this, developing an automated system to detect and classify harmful content is essential. This study develops a deep learning model using Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) to detect SARA-related comments on Twitter. The method involves data collection through web scraping, followed by cleaning, manual labeling, and text preprocessing. To address data imbalance, SMOTE (Synthetic Minority Over-sampling Technique) is applied, while early stopping prevents overfitting. Model performance is evaluated using precision, recall, and F1-score. The results demonstrate that SMOTE significantly improves model performance, particularly in detecting minority-class SARA comments. CNN+SMOTE achieves a accuracy of 93%, and BiLSTM+SMOTE records a recall of 88%, effectively capturing patterns in SARA and non-SARA data. With SMOTE and early stopping, the model successfully manages class imbalance and reduces overfitting. This research supports efforts to curtail hate speech on social media, especially in the Indonesian context, where SARA-related issues often dominate public discourse.

Keywords

• Deep Learning
• SARA Comments
• SARA Detection
• SMOTE
• Social Media Classification