earticle

영어

The proliferation of social media has facilitated communication but has also accelerated the spread of hate speech and offensive language. The automatic detection of such harmful content is essential for ensuring the safety and integrity of online platforms. This study proposes a practical classification system for hate speech detection using Google’s BERT (Bidirectional Encoder Representations from Transformers) model. A dataset consisting of 24,783 tweets was utilized, categorized into three labels: Hate Speech, Offensive Language, and Neither. To address the issue of data imbalance, techniques such as SMOTE, class weighting, and threshold optimization were applied. The preprocessing pipeline included the removal of URLs, user tags, and special characters to enhance data consistency. Compared with traditional machine learning models—Logistic Regression, Random Forest, and SVM—the BERT-based model demonstrated superior performance in both contextual understanding and classification accuracy. Model performance was evaluated using confusion matrices, precision, recall, F1-score, as well as ROC and Precision-Recall curves, showing particularly strong results for the Offensive Language category. Future research will extend this work by applying advanced transformer-based models such as RoBERTa and GPT-3, and by constructing multilingual datasets to develop a scalable, real-time detection system applicable to global social media platforms.