AN ENSEMBLE CLUSTERING-BASED APPROACH fOR EMOTION DETECTION IN YOUTUBE COMMENTS: A CASE STUDY ON THE GAZA WAR

Abstract

Emotion detection from text has become a critical research direction in natural language pro- cessing, particularly with the rise of emotionally charged and multilingual content on platforms like YouTube. This study focuses on unsupervised emotion detection from short YouTube comments related to the Gaza war, using data in both English and Arabic. It explores the role of emojis and different levels of textual "representation at the word and sentence levels" in enhancing emotional interpretation. By employing ensemble clustering techniques, including co-association matrices and the Multi-Metric Genetic Algorithm (MM-GA), the research aims to uncover latent emotional structures without relying on pre-labeled data. The results demonstrate that sentence-level representations, especially when combined with emojis, significantly outperform word-level representations in terms of clustering quality, achiev- ing higher scores on evaluation metrics such as Silhouette, Calinski-Harabasz, and Davies-Bouldin indices. Emojis emerge as strong emotional indicators, In particular, the Silhouette Score reached 0.3367 for Arabic comments, while for English comments, the score reached 0.3130, indicating strong cluster separation and cohesion. Emojis emerge as strong emotional indicators, particularly in Arabic comments, which often exhibit linguistic ambiguity. Comparative analysis shows that MM-GA and co-association-based clustering methods provide more stable and coherent perfor- mance than majority voting or traditional Mirkin-based approaches. Overall, the findings confirm the effectiveness of ensemble clustering in detecting emotional patterns from unstructured and di- verse YouTube data. The proposed framework offers a scalable and language-aware solution for unsupervised emotion detection, particularly suited to multilingual and politically sensitive con- texts