10th International Congress on Information and Communication Technology

Authors - Liyana Safra Zaabar, Noor Afiza Mat Razali, Sharifah Nabila S Azli Sham, Fazlina Mohd Ali
Abstract - The increasing spread of textual content on social media, driven by the rise of Large Language Models (LLMs), has highlighted the importance of sentiment analysis in detecting threats, racial abuse, violence, and implied warnings. The subtlety and ambiguity of language present challenges in developing effective frameworks for threat detection, particularly within the political security domain. While significant research has explored hate speech and offensive content, few studies focus on detecting threats using sentiment analysis in this context. Leveraging advancements in Natural Language Processing (NLP), this study employs the NRC Emotion Lexicon to label emotions in a political-domain social media dataset. The Bidirectional Encoder Representations from Transformers (BERT) model was applied to improve threat detection accuracy. The proposed model achieved an AUC value of 87%, with the BERT model achieving 91% accuracy, 90.5% precision, 81.3% recall and F1-score of 91%, outperforming baseline models. These findings demonstrate the effectiveness of sentiment and emotion-based features in improving threat detection accuracy, providing a robust framework for political security applications.