A New Framework for Nuanced Sentiment Analysis in Political Communication: A Machine Learning Approach to the Case Study of His Majesty King Abdullah II's UN Speech

Sentiment analysis has emerged as a crucial tool for understanding public opinion across various fields, yet studies focusing specifically on political speeches remain limited. This research employs advanced computational techniques to analyze the sentiments expressed in King Abdullah II's speech at the United Nations, which addresses pressing humanitarian concerns amid the ongoing crisis in Gaza. While traditional sentiment analysis typically classifies sentiments into positive, negative, and neutral categories, this study refines this approach by introducing five nuanced classifications- Empathetic, Call to Action, Critique of Power, Reassuring, and Urgent. Utilizing state-of-the-art natural language processing methodologies, we directly analyze the speech text, employing various supervised machine-learning algorithms to assess the effectiveness of these models. Preliminary findings reveal an accuracy rate of approximately 85%, with the Support Vector Machine algorithm demonstrating exceptional performance across all sentiment categories. Notably, The analysis uncovers a predominance of empathetic and critical sentiments, underscoring the King's deep concern for humanitarian issues and the urgent need for action. This study highlights the classifier performance through metrics such as accuracy, precision, recall, and F1-score and fills a significant gap in the literature by providing patterns into political communication strategies during crises. The findings offer valuable implications for policymakers and community leaders, enabling them to foster civic engagement and inspire citizens to contribute to national development. By enhancing our understanding of the emotional nuances within political discourse, this research serves as a model for future sentiment analysis studies, ultimately contributing to the resilience of communities in the Arab region. Limitations include the single-speaker, single-speech dataset; we discuss directions for broader validation on larger, multi-speaker corpora.